sync with parent

* feat: add calendar event data to transcript webhook payload and implement get_by_id method

* Update server/reflector/worker/webhook.py

Co-authored-by: pr-agent-monadical[bot] <198624643+pr-agent-monadical[bot]@users.noreply.github.com>

* Update server/reflector/worker/webhook.py

Co-authored-by: pr-agent-monadical[bot] <198624643+pr-agent-monadical[bot]@users.noreply.github.com>

* style: format conditional time fields with line breaks for better readability

* docs: add calendar event fields to transcript.completed webhook payload schema

---------

Co-authored-by: pr-agent-monadical[bot] <198624643+pr-agent-monadical[bot]@users.noreply.github.com>

.github/workflows/db_migrations.yml

@@ -2,6 +2,8 @@ name: Test Database Migrations
 
 on:
   push:
+    branches:
+      - main
     paths:
       - "server/migrations/**"
       - "server/reflector/db/**"
@@ -17,6 +19,9 @@ on:
 jobs:
   test-migrations:
     runs-on: ubuntu-latest
+    concurrency:
+      group: db-ubuntu-latest-${{ github.ref }}
+      cancel-in-progress: true
     services:
       postgres:
         image: postgres:17

.github/workflows/docker-frontend.yml

@@ -0,0 +1,57 @@
+name: Build and Push Frontend Docker Image
+
+on:
+  push:
+    branches:
+      - main
+    paths:
+      - 'www/**'
+      - '.github/workflows/docker-frontend.yml'
+  workflow_dispatch:
+
+env:
+  REGISTRY: ghcr.io
+  IMAGE_NAME: ${{ github.repository }}-frontend
+
+jobs:
+  build-and-push:
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
+      packages: write
+
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+
+      - name: Log in to GitHub Container Registry
+        uses: docker/login-action@v3
+        with:
+          registry: ${{ env.REGISTRY }}
+          username: ${{ github.actor }}
+          password: ${{ secrets.GITHUB_TOKEN }}
+
+      - name: Extract metadata
+        id: meta
+        uses: docker/metadata-action@v5
+        with:
+          images: ${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}
+          tags: |
+            type=ref,event=branch
+            type=sha,prefix={{branch}}-
+            type=raw,value=latest,enable={{is_default_branch}}
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+
+      - name: Build and push Docker image
+        uses: docker/build-push-action@v5
+        with:
+          context: ./www
+          file: ./www/Dockerfile
+          push: true
+          tags: ${{ steps.meta.outputs.tags }}
+          labels: ${{ steps.meta.outputs.labels }}
+          cache-from: type=gha
+          cache-to: type=gha,mode=max
+          platforms: linux/amd64,linux/arm64

.github/workflows/test_next_server.yml

@@ -0,0 +1,45 @@
+name: Test Next Server
+
+on:
+  pull_request:
+    paths:
+      - "www/**"
+  push:
+    branches:
+      - main
+    paths:
+      - "www/**"
+
+jobs:
+  test-next-server:
+    runs-on: ubuntu-latest
+
+    defaults:
+      run:
+        working-directory: ./www
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Setup Node.js
+      uses: actions/setup-node@v4
+      with:
+        node-version: '20'
+
+    - name: Install pnpm
+      uses: pnpm/action-setup@v4
+      with:
+        version: 8
+
+    - name: Setup Node.js cache
+      uses: actions/setup-node@v4
+      with:
+        node-version: '20'
+        cache: 'pnpm'
+        cache-dependency-path: './www/pnpm-lock.yaml'
+
+    - name: Install dependencies
+      run: pnpm install
+
+    - name: Run tests
+      run: pnpm test

.github/workflows/test_server.yml

@@ -5,12 +5,17 @@ on:
     paths:
       - "server/**"
   push:
+    branches:
+      - main
     paths:
       - "server/**"
 
 jobs:
   pytest:
     runs-on: ubuntu-latest
+    concurrency:
+      group: pytest-${{ github.ref }}
+      cancel-in-progress: true
     services:
       redis:
         image: redis:6
@@ -30,6 +35,9 @@ jobs:
 
   docker-amd64:
     runs-on: linux-amd64
+    concurrency:
+      group: docker-amd64-${{ github.ref }}
+      cancel-in-progress: true
     steps:
       - uses: actions/checkout@v4
       - name: Set up Docker Buildx
@@ -45,6 +53,9 @@ jobs:
 
   docker-arm64:
     runs-on: linux-arm64
+    concurrency:
+      group: docker-arm64-${{ github.ref }}
+      cancel-in-progress: true
     steps:
       - uses: actions/checkout@v4
       - name: Set up Docker Buildx

.gitignore

@@ -14,4 +14,7 @@ data/
 www/REFACTOR.md
 www/reload-frontend
 server/test.sqlite
-CLAUDE.local.md
+CLAUDE.local.md
+www/.env.development
+www/.env.production
+.playwright-mcp

.gitleaksignore

@@ -0,0 +1 @@
+b9d891d3424f371642cb032ecfd0e2564470a72c:server/tests/test_transcripts_recording_deletion.py:generic-api-key:15

.pre-commit-config.yaml

@@ -27,3 +27,8 @@ repos:
         files: ^server/
       - id: ruff-format
         files: ^server/
+
+  - repo: https://github.com/gitleaks/gitleaks
+    rev: v8.28.0
+    hooks:
+      - id: gitleaks

CHANGELOG.md

@@ -1,5 +1,151 @@
 # Changelog
 
+## [0.14.0](https://github.com/Monadical-SAS/reflector/compare/v0.13.1...v0.14.0) (2025-10-08)
+
+
+### Features
+
+* Add calendar event data to transcript webhook payload ([#689](https://github.com/Monadical-SAS/reflector/issues/689)) ([5f6910e](https://github.com/Monadical-SAS/reflector/commit/5f6910e5131b7f28f86c9ecdcc57fed8412ee3cd))
+* container build for www / github ([#672](https://github.com/Monadical-SAS/reflector/issues/672)) ([969bd84](https://github.com/Monadical-SAS/reflector/commit/969bd84fcc14851d1a101412a0ba115f1b7cde82))
+* docker-compose for production frontend ([#664](https://github.com/Monadical-SAS/reflector/issues/664)) ([5bf64b5](https://github.com/Monadical-SAS/reflector/commit/5bf64b5a41f64535e22849b4bb11734d4dbb4aae))
+
+
+### Bug Fixes
+
+* restore feature boolean logic ([#671](https://github.com/Monadical-SAS/reflector/issues/671)) ([3660884](https://github.com/Monadical-SAS/reflector/commit/36608849ec64e953e3be456172502762e3c33df9))
+* security review ([#656](https://github.com/Monadical-SAS/reflector/issues/656)) ([5d98754](https://github.com/Monadical-SAS/reflector/commit/5d98754305c6c540dd194dda268544f6d88bfaf8))
+* update transcript list on reprocess ([#676](https://github.com/Monadical-SAS/reflector/issues/676)) ([9a71af1](https://github.com/Monadical-SAS/reflector/commit/9a71af145ee9b833078c78d0c684590ab12e9f0e))
+* upgrade nemo toolkit ([#678](https://github.com/Monadical-SAS/reflector/issues/678)) ([eef6dc3](https://github.com/Monadical-SAS/reflector/commit/eef6dc39037329b65804297786d852dddb0557f9))
+
+## [0.13.1](https://github.com/Monadical-SAS/reflector/compare/v0.13.0...v0.13.1) (2025-09-22)
+
+
+### Bug Fixes
+
+* TypeError on not all arguments converted during string formatting in logger ([#667](https://github.com/Monadical-SAS/reflector/issues/667)) ([565a629](https://github.com/Monadical-SAS/reflector/commit/565a62900f5a02fc946b68f9269a42190ed70ab6))
+
+## [0.13.0](https://github.com/Monadical-SAS/reflector/compare/v0.12.1...v0.13.0) (2025-09-19)
+
+
+### Features
+
+* room form edit with enter ([#662](https://github.com/Monadical-SAS/reflector/issues/662)) ([47716f6](https://github.com/Monadical-SAS/reflector/commit/47716f6e5ddee952609d2fa0ffabdfa865286796))
+
+
+### Bug Fixes
+
+* invalid cleanup call ([#660](https://github.com/Monadical-SAS/reflector/issues/660)) ([0abcebf](https://github.com/Monadical-SAS/reflector/commit/0abcebfc9491f87f605f21faa3e53996fafedd9a))
+
+## [0.12.1](https://github.com/Monadical-SAS/reflector/compare/v0.12.0...v0.12.1) (2025-09-17)
+
+
+### Bug Fixes
+
+* production blocked because having existing meeting with room_id null ([#657](https://github.com/Monadical-SAS/reflector/issues/657)) ([870e860](https://github.com/Monadical-SAS/reflector/commit/870e8605171a27155a9cbee215eeccb9a8d6c0a2))
+
+## [0.12.0](https://github.com/Monadical-SAS/reflector/compare/v0.11.0...v0.12.0) (2025-09-17)
+
+
+### Features
+
+* calendar integration ([#608](https://github.com/Monadical-SAS/reflector/issues/608)) ([6f680b5](https://github.com/Monadical-SAS/reflector/commit/6f680b57954c688882c4ed49f40f161c52a00a24))
+* self-hosted gpu api ([#636](https://github.com/Monadical-SAS/reflector/issues/636)) ([ab859d6](https://github.com/Monadical-SAS/reflector/commit/ab859d65a6bded904133a163a081a651b3938d42))
+
+
+### Bug Fixes
+
+* ignore player hotkeys for text inputs ([#646](https://github.com/Monadical-SAS/reflector/issues/646)) ([fa049e8](https://github.com/Monadical-SAS/reflector/commit/fa049e8d068190ce7ea015fd9fcccb8543f54a3f))
+
+## [0.11.0](https://github.com/Monadical-SAS/reflector/compare/v0.10.0...v0.11.0) (2025-09-16)
+
+
+### Features
+
+* remove profanity filter that was there for conference ([#652](https://github.com/Monadical-SAS/reflector/issues/652)) ([b42f7cf](https://github.com/Monadical-SAS/reflector/commit/b42f7cfc606783afcee792590efcc78b507468ab))
+
+
+### Bug Fixes
+
+* zulip and consent handler on the file pipeline ([#645](https://github.com/Monadical-SAS/reflector/issues/645)) ([5f143fe](https://github.com/Monadical-SAS/reflector/commit/5f143fe3640875dcb56c26694254a93189281d17))
+* zulip stream and topic selection in share dialog ([#644](https://github.com/Monadical-SAS/reflector/issues/644)) ([c546e69](https://github.com/Monadical-SAS/reflector/commit/c546e69739e68bb74fbc877eb62609928e5b8de6))
+
+## [0.10.0](https://github.com/Monadical-SAS/reflector/compare/v0.9.0...v0.10.0) (2025-09-11)
+
+
+### Features
+
+* replace nextjs-config with environment variables ([#632](https://github.com/Monadical-SAS/reflector/issues/632)) ([369ecdf](https://github.com/Monadical-SAS/reflector/commit/369ecdff13f3862d926a9c0b87df52c9d94c4dde))
+
+
+### Bug Fixes
+
+* anonymous users transcript permissions ([#621](https://github.com/Monadical-SAS/reflector/issues/621)) ([f81fe99](https://github.com/Monadical-SAS/reflector/commit/f81fe9948a9237b3e0001b2d8ca84f54d76878f9))
+* auth post ([#624](https://github.com/Monadical-SAS/reflector/issues/624)) ([cde99ca](https://github.com/Monadical-SAS/reflector/commit/cde99ca2716f84ba26798f289047732f0448742e))
+* auth post ([#626](https://github.com/Monadical-SAS/reflector/issues/626)) ([3b85ff3](https://github.com/Monadical-SAS/reflector/commit/3b85ff3bdf4fb053b103070646811bc990c0e70a))
+* auth post ([#627](https://github.com/Monadical-SAS/reflector/issues/627)) ([962038e](https://github.com/Monadical-SAS/reflector/commit/962038ee3f2a555dc3c03856be0e4409456e0996))
+* missing follow_redirects=True on modal endpoint ([#630](https://github.com/Monadical-SAS/reflector/issues/630)) ([fc363bd](https://github.com/Monadical-SAS/reflector/commit/fc363bd49b17b075e64f9186e5e0185abc325ea7))
+* sync backend and frontend token refresh logic ([#614](https://github.com/Monadical-SAS/reflector/issues/614)) ([5a5b323](https://github.com/Monadical-SAS/reflector/commit/5a5b3233820df9536da75e87ce6184a983d4713a))
+
+## [0.9.0](https://github.com/Monadical-SAS/reflector/compare/v0.8.2...v0.9.0) (2025-09-06)
+
+
+### Features
+
+* frontend openapi react query ([#606](https://github.com/Monadical-SAS/reflector/issues/606)) ([c4d2825](https://github.com/Monadical-SAS/reflector/commit/c4d2825c81f81ad8835629fbf6ea8c7383f8c31b))
+
+
+### Bug Fixes
+
+* align whisper transcriber api with parakeet ([#602](https://github.com/Monadical-SAS/reflector/issues/602)) ([0663700](https://github.com/Monadical-SAS/reflector/commit/0663700a615a4af69a03c96c410f049e23ec9443))
+* kv use tls explicit ([#610](https://github.com/Monadical-SAS/reflector/issues/610)) ([08d88ec](https://github.com/Monadical-SAS/reflector/commit/08d88ec349f38b0d13e0fa4cb73486c8dfd31836))
+* source kind for file processing ([#601](https://github.com/Monadical-SAS/reflector/issues/601)) ([dc82f8b](https://github.com/Monadical-SAS/reflector/commit/dc82f8bb3bdf3ab3d4088e592a30fd63907319e1))
+* token refresh locking ([#613](https://github.com/Monadical-SAS/reflector/issues/613)) ([7f5a4c9](https://github.com/Monadical-SAS/reflector/commit/7f5a4c9ddc7fd098860c8bdda2ca3b57f63ded2f))
+
+## [0.8.2](https://github.com/Monadical-SAS/reflector/compare/v0.8.1...v0.8.2) (2025-08-29)
+
+
+### Bug Fixes
+
+* search-logspam ([#593](https://github.com/Monadical-SAS/reflector/issues/593)) ([695d1a9](https://github.com/Monadical-SAS/reflector/commit/695d1a957d4cd862753049f9beed88836cabd5ab))
+
+## [0.8.1](https://github.com/Monadical-SAS/reflector/compare/v0.8.0...v0.8.1) (2025-08-29)
+
+
+### Bug Fixes
+
+* make webhook secret/url allowing null ([#590](https://github.com/Monadical-SAS/reflector/issues/590)) ([84a3812](https://github.com/Monadical-SAS/reflector/commit/84a381220bc606231d08d6f71d4babc818fa3c75))
+
+## [0.8.0](https://github.com/Monadical-SAS/reflector/compare/v0.7.3...v0.8.0) (2025-08-29)
+
+
+### Features
+
+* **cleanup:** add automatic data retention for public instances ([#574](https://github.com/Monadical-SAS/reflector/issues/574)) ([6f0c7c1](https://github.com/Monadical-SAS/reflector/commit/6f0c7c1a5e751713366886c8e764c2009e12ba72))
+* **rooms:** add webhook for transcript completion ([#578](https://github.com/Monadical-SAS/reflector/issues/578)) ([88ed7cf](https://github.com/Monadical-SAS/reflector/commit/88ed7cfa7804794b9b54cad4c3facc8a98cf85fd))
+
+
+### Bug Fixes
+
+* file pipeline status reporting and websocket updates ([#589](https://github.com/Monadical-SAS/reflector/issues/589)) ([9dfd769](https://github.com/Monadical-SAS/reflector/commit/9dfd76996f851cc52be54feea078adbc0816dc57))
+* Igor/evaluation ([#575](https://github.com/Monadical-SAS/reflector/issues/575)) ([124ce03](https://github.com/Monadical-SAS/reflector/commit/124ce03bf86044c18313d27228a25da4bc20c9c5))
+* optimize parakeet transcription batching algorithm ([#577](https://github.com/Monadical-SAS/reflector/issues/577)) ([7030e0f](https://github.com/Monadical-SAS/reflector/commit/7030e0f23649a8cf6c1eb6d5889684a41ce849ec))
+
+## [0.7.3](https://github.com/Monadical-SAS/reflector/compare/v0.7.2...v0.7.3) (2025-08-22)
+
+
+### Bug Fixes
+
+* cleaned repo, and get git-leaks clean ([359280d](https://github.com/Monadical-SAS/reflector/commit/359280dd340433ba4402ed69034094884c825e67))
+* restore previous behavior on live pipeline + audio downscaler ([#561](https://github.com/Monadical-SAS/reflector/issues/561)) ([9265d20](https://github.com/Monadical-SAS/reflector/commit/9265d201b590d23c628c5f19251b70f473859043))
+
+## [0.7.2](https://github.com/Monadical-SAS/reflector/compare/v0.7.1...v0.7.2) (2025-08-21)
+
+
+### Bug Fixes
+
+* docker image not loading libgomp.so.1 for torch ([#560](https://github.com/Monadical-SAS/reflector/issues/560)) ([773fccd](https://github.com/Monadical-SAS/reflector/commit/773fccd93e887c3493abc2e4a4864dddce610177))
+* include shared rooms to search ([#558](https://github.com/Monadical-SAS/reflector/issues/558)) ([499eced](https://github.com/Monadical-SAS/reflector/commit/499eced3360b84fb3a90e1c8a3b554290d21adc2))
+
 ## [0.7.1](https://github.com/Monadical-SAS/reflector/compare/v0.7.0...v0.7.1) (2025-08-21)
 
 

CLAUDE.md

@@ -66,7 +66,6 @@ pnpm install
 
 # Copy configuration templates
 cp .env_template .env
-cp config-template.ts config.ts
 ```
 
 **Development:**
@@ -152,7 +151,7 @@ All endpoints prefixed `/v1/`:
 
 **Frontend** (`www/.env`):
 - `NEXTAUTH_URL`, `NEXTAUTH_SECRET` - Authentication configuration
-- `NEXT_PUBLIC_REFLECTOR_API_URL` - Backend API endpoint
+- `REFLECTOR_API_URL` - Backend API endpoint
 - `REFLECTOR_DOMAIN_CONFIG` - Feature flags and domain settings
 
 ## Testing Strategy

CODER_BRIEFING.md

@@ -0,0 +1,345 @@
+# Multi-Provider Video Platform Implementation - Coder Briefing
+
+## Your Mission
+
+Implement multi-provider video platform support in Reflector, allowing the system to work with both Whereby and Daily.co video conferencing providers. The goal is to abstract the current Whereby-only implementation and add Daily.co as a second provider, with the ability to switch between them via environment variables.
+
+**Branch:** `igor/dailico-2` (you're already on it)
+
+**Estimated Time:** 12-16 hours (senior engineer)
+
+**Complexity:** Medium-High (requires careful integration with existing codebase)
+
+---
+
+## What You Have
+
+### 1. **PLAN.md** - Your Technical Specification (2,452 lines)
+   - Complete step-by-step implementation guide
+   - All code examples you need
+   - Architecture diagrams and design rationale
+   - Testing strategy and success metrics
+   - **Read this first** to understand the overall approach
+
+### 2. **IMPLEMENTATION_GUIDE.md** - Your Practical Guide
+   - What to copy vs. adapt vs. rewrite
+   - Common pitfalls and how to avoid them
+   - Verification checklists for each phase
+   - Decision trees for implementation choices
+   - **Use this as your day-to-day reference**
+
+### 3. **Reference Implementation** - `./reflector-dailyco-reference/`
+   - Working implementation from 2.5 months ago
+   - Good architecture and patterns
+   - **BUT:** 91 commits behind current main, DO NOT merge directly
+   - Use for inspiration and code patterns only
+
+---
+
+## Critical Context: Why Not Just Merge?
+
+The reference branch (`origin/igor/feat-dailyco`) was started on August 1, 2025 and is now severely diverged from main:
+
+- **91 commits behind main**
+- Main has 12x more changes (45,840 insertions vs 3,689)
+- Main added: calendar integration, webhooks, full-text search, React Query migration, security fixes
+- Reference removed: features that main still has and needs
+
+**Merging would be a disaster.** We're implementing fresh on current main, using the reference for validated patterns.
+
+---
+
+## High-Level Approach
+
+### Phase 1: Analysis (2 hours)
+- Study current Whereby integration
+- Define abstraction requirements
+- Create standard data models
+
+### Phase 2: Abstraction Layer (4-5 hours)
+- Build platform abstraction (base class, registry, factory)
+- Extract Whereby into the abstraction
+- Update database schema (add `platform` field)
+- Integrate into rooms.py **without breaking calendar/webhooks**
+
+### Phase 3: Daily.co Implementation (4-5 hours)
+- Implement Daily.co client
+- Add webhook handler
+- Create frontend components (rewrite API calls for React Query)
+- Add recording processing
+
+### Phase 4: Testing (2-3 hours)
+- Unit tests for platform abstraction
+- Integration tests for webhooks
+- Manual testing with both providers
+
+---
+
+## Key Files You'll Touch
+
+### Backend (New)
+```
+server/reflector/video_platforms/
+├── __init__.py
+├── base.py              ← Abstract base class
+├── models.py            ← Platform, MeetingData, VideoPlatformConfig
+├── registry.py          ← Platform registration system
+├── factory.py           ← Client creation and config
+├── whereby.py           ← Whereby client wrapper
+├── daily.py             ← Daily.co client
+└── mock.py              ← Mock client for testing
+
+server/reflector/views/daily.py       ← Daily.co webhooks
+server/tests/test_video_platforms.py  ← Platform tests
+server/tests/test_daily_webhook.py    ← Webhook tests
+```
+
+### Backend (Modified - Careful!)
+```
+server/reflector/settings.py          ← Add Daily.co settings
+server/reflector/db/rooms.py          ← Add platform field, PRESERVE calendar fields
+server/reflector/db/meetings.py       ← Add platform field
+server/reflector/views/rooms.py       ← Integrate abstraction, PRESERVE calendar/webhooks
+server/reflector/worker/process.py    ← Add process_recording_from_url task
+server/reflector/app.py               ← Register daily router
+server/env.example                    ← Document new env vars
+```
+
+### Frontend (New)
+```
+www/app/[roomName]/components/
+├── RoomContainer.tsx    ← Platform router
+├── DailyRoom.tsx        ← Daily.co component (rewrite API calls!)
+└── WherebyRoom.tsx      ← Extract existing logic
+```
+
+### Frontend (Modified)
+```
+www/app/[roomName]/page.tsx           ← Use RoomContainer
+www/package.json                      ← Add @daily-co/daily-js
+```
+
+### Database
+```
+server/migrations/versions/XXXXXX_add_platform_support.py  ← Generate fresh migration
+```
+
+---
+
+## Critical Warnings ⚠️
+
+### 1. **DO NOT Copy Database Migrations**
+The reference migration has the wrong `down_revision` and is based on old schema.
+```bash
+# Instead:
+cd server
+uv run alembic revision -m "add_platform_support"
+# Then edit the generated file
+```
+
+### 2. **DO NOT Remove Main's Features**
+Main has calendar integration, webhooks, ICS sync that reference doesn't have.
+When modifying `rooms.py`, only change meeting creation logic, preserve everything else.
+
+### 3. **DO NOT Copy Frontend API Calls**
+Reference uses old OpenAPI client. Main uses React Query.
+Check how main currently makes API calls and replicate that pattern.
+
+### 4. **DO NOT Copy package.json/migrations**
+These files are severely outdated in reference.
+
+### 5. **Preserve Type Safety**
+Use `TYPE_CHECKING` imports to avoid circular dependencies:
+```python
+from typing import TYPE_CHECKING
+if TYPE_CHECKING:
+    from reflector.db.rooms import Room
+```
+
+---
+
+## How to Start
+
+### Day 1 Morning: Setup & Understanding (2-3 hours)
+```bash
+# 1. Verify you're on the right branch
+git branch
+# Should show: igor/dailico-2
+
+# 2. Read the docs (in order)
+# - PLAN.md (skim to understand scope, read Phase 1 carefully)
+# - IMPLEMENTATION_GUIDE.md (read fully, bookmark it)
+
+# 3. Study current Whereby integration
+cat server/reflector/views/rooms.py | grep -A 20 "whereby"
+cat www/app/[roomName]/page.tsx
+
+# 4. Check reference implementation structure
+ls -la reflector-dailyco-reference/server/reflector/video_platforms/
+```
+
+### Day 1 Afternoon: Phase 1 Execution (2-3 hours)
+```bash
+# 5. Copy video_platforms directory from reference
+cp -r reflector-dailyco-reference/server/reflector/video_platforms/ \
+      server/reflector/
+
+# 6. Review and fix imports
+cd server
+uv run ruff check reflector/video_platforms/
+
+# 7. Add settings to settings.py (see PLAN.md Phase 2.7)
+
+# 8. Test imports work
+uv run python -c "from reflector.video_platforms import create_platform_client; print('OK')"
+```
+
+### Day 2: Phase 2 - Database & Integration (4-5 hours)
+```bash
+# 9. Generate migration
+uv run alembic revision -m "add_platform_support"
+# Edit the file following PLAN.md Phase 2.8
+
+# 10. Update Room/Meeting models
+# Add platform field, PRESERVE all existing fields
+
+# 11. Integrate into rooms.py
+# Carefully modify meeting creation, preserve calendar/webhooks
+
+# 12. Add Daily.co webhook handler
+cp reflector-dailyco-reference/server/reflector/views/daily.py \
+   server/reflector/views/
+# Register in app.py
+```
+
+### Day 3: Phase 3 - Frontend & Testing (4-5 hours)
+```bash
+# 13. Create frontend components
+mkdir -p www/app/[roomName]/components
+
+# 14. Add Daily.co dependency
+cd www
+pnpm add @daily-co/daily-js@^0.81.0
+
+# 15. Create RoomContainer, DailyRoom, WherebyRoom
+# IMPORTANT: Rewrite API calls using React Query patterns
+
+# 16. Regenerate types
+pnpm openapi
+
+# 17. Copy and adapt tests
+cp reflector-dailyco-reference/server/tests/test_*.py server/tests/
+
+# 18. Run tests
+cd server
+REDIS_HOST=localhost \
+CELERY_BROKER_URL=redis://localhost:6379/1 \
+uv run pytest tests/test_video_platforms.py -v
+```
+
+---
+
+## Verification Checklist
+
+After implementation, all of these must pass:
+
+**Backend:**
+- [ ] `cd server && uv run ruff check .` passes
+- [ ] `uv run alembic upgrade head` works cleanly
+- [ ] `uv run pytest tests/test_video_platforms.py` passes
+- [ ] Can import: `from reflector.video_platforms import create_platform_client`
+- [ ] Settings has all Daily.co variables
+
+**Frontend:**
+- [ ] `cd www && pnpm lint` passes
+- [ ] No TypeScript errors
+- [ ] `pnpm openapi` generates platform field
+- [ ] No `@ts-ignore` for platform field
+
+**Integration:**
+- [ ] Whereby meetings still work (existing flow unchanged)
+- [ ] Calendar/webhook features still work in rooms.py
+- [ ] env.example documents all new variables
+
+---
+
+## When You're Stuck
+
+### Check These Resources:
+1. **PLAN.md** - Detailed code examples for your exact scenario
+2. **IMPLEMENTATION_GUIDE.md** - Common pitfalls section
+3. **Reference code** - See how it was solved before
+4. **Git diff** - Compare reference to your implementation
+
+### Compare Files:
+```bash
+# See what reference did
+diff reflector-dailyco-reference/server/reflector/views/rooms.py \
+     server/reflector/views/rooms.py
+
+# See what changed in main since reference branch
+git log --oneline --since="2025-08-01" -- server/reflector/views/rooms.py
+```
+
+### Common Issues:
+- **Circular imports:** Use `TYPE_CHECKING` pattern
+- **Tests fail with postgres error:** Use `REDIS_HOST=localhost` env vars
+- **Frontend API calls broken:** Check current React Query patterns in main
+- **Migrations fail:** Ensure you generated fresh, not copied
+
+---
+
+## Success Looks Like
+
+When you're done:
+- ✅ All tests pass
+- ✅ Linting passes
+- ✅ Can create Whereby meetings (unchanged behavior)
+- ✅ Can create Daily.co meetings (with env vars)
+- ✅ Calendar/webhooks still work
+- ✅ Frontend has no TypeScript errors
+- ✅ Platform selection via environment variables works
+
+---
+
+## Communication
+
+If you need clarification on requirements, have questions about architecture decisions, or find issues with the spec, document them clearly with:
+- What you expected
+- What you found
+- Your proposed solution
+
+The PLAN.md document is comprehensive but you may find edge cases. Use your engineering judgment and document decisions.
+
+---
+
+## Final Notes
+
+**This is not a simple copy-paste job.** You're doing careful integration work where you need to:
+- Understand the abstraction pattern (PLAN.md)
+- Preserve all of main's features
+- Adapt reference code to current patterns
+- Think about edge cases and testing
+
+Take your time with Phase 2 (rooms.py integration) - that's where most bugs will come from if you accidentally break calendar/webhook features.
+
+**Good luck! You've got comprehensive specs, working reference code, and a clean starting point. You can do this.**
+
+---
+
+## Quick Reference
+
+```bash
+# Your workspace
+├── PLAN.md                        ← Complete technical spec (read first)
+├── IMPLEMENTATION_GUIDE.md        ← Practical guide (bookmark this)
+├── CODER_BRIEFING.md             ← This file
+└── reflector-dailyco-reference/   ← Reference implementation (inspiration only)
+
+# Key commands
+cd server && uv run ruff check .                    # Lint backend
+cd www && pnpm lint                                  # Lint frontend
+cd server && uv run alembic revision -m "..."       # Create migration
+cd www && pnpm openapi                              # Regenerate types
+cd server && uv run pytest -v                       # Run tests
+```

IMPLEMENTATION_GUIDE.md

@@ -0,0 +1,489 @@
+# Daily.co Implementation Guide
+
+## Overview
+Implement multi-provider video platform support (Whereby + Daily.co) following PLAN.md.
+
+## Reference Code Location
+- **Reference branch:** `origin/igor/feat-dailyco` (on remote)
+- **Worktree location:** `./reflector-dailyco-reference/`
+- **Status:** Reference only - DO NOT merge or copy directly
+
+## What Exists in Reference Branch (For Inspiration)
+
+### ✅ Can Use As Reference (Well-Implemented)
+```
+server/reflector/video_platforms/
+├── base.py              ← Platform abstraction (good design, copy-safe)
+├── models.py            ← Data models (copy-safe)
+├── registry.py          ← Registry pattern (copy-safe)
+├── factory.py           ← Factory pattern (needs settings updates)
+├── whereby.py           ← Whereby client (needs adaptation)
+├── daily.py             ← Daily.co client (needs adaptation)
+└── mock.py              ← Mock client (copy-safe for tests)
+
+server/reflector/views/daily.py       ← Webhook handler (needs adaptation)
+server/tests/test_video_platforms.py  ← Tests (good reference)
+server/tests/test_daily_webhook.py    ← Tests (good reference)
+
+www/app/[roomName]/components/
+├── RoomContainer.tsx    ← Platform router (needs React Query)
+├── DailyRoom.tsx        ← Daily component (needs React Query)
+└── WherebyRoom.tsx      ← Whereby extraction (needs React Query)
+```
+
+### ⚠️ Needs Significant Changes (Use Logic Only)
+- `server/reflector/db/rooms.py` - Reference removed calendar/webhook fields that main has
+- `server/reflector/db/meetings.py` - Same issue (missing user_id handling differences)
+- `server/reflector/views/rooms.py` - Main has calendar integration, webhooks, ICS sync
+- `server/reflector/worker/process.py` - Main has different recording flow
+- Migration files - Must regenerate against current main schema
+
+### ❌ Do NOT Use (Outdated/Incompatible)
+- `package.json`/`pnpm-lock.yaml` - Main uses different dependency versions
+- Frontend API client calls - Main uses React Query (reference uses old OpenAPI client)
+- Database migrations - Must create new ones from scratch
+- Any files that delete features present in main (search, calendar, webhooks)
+
+## Key Differences: Reference vs Current Main
+
+| Aspect | Reference Branch | Current Main | Action Required |
+|--------|------------------|--------------|-----------------|
+| **API client** | Old OpenAPI generated | React Query hooks | Rewrite all API calls |
+| **Database schema** | Simplified (removed features) | Has calendar, webhooks, full-text search | Merge carefully, preserve main features |
+| **Settings** | Aug 2025 structure | Current structure | Adapt carefully |
+| **Migrations** | Branched from Aug 1 | Current main (91+ commits ahead) | Regenerate from scratch |
+| **Frontend deps** | `@daily-co/daily-js@0.81.0` | Check current versions | Update to compatible versions |
+| **Package manager** | yarn | pnpm (maybe both?) | Use what main uses |
+
+## Branch Divergence Analysis
+
+**The reference branch is 91 commits behind main and severely diverged:**
+- Reference: 8 commits, 3,689 insertions, 425 deletions
+- Main since divergence: 320 files changed, 45,840 insertions, 16,827 deletions
+- **Main has 12x more changes**
+
+**Major features in main that reference lacks:**
+1. Calendar integration (ICS sync with rooms)
+2. Self-hosted GPU API infrastructure
+3. Frontend OpenAPI React Query migration
+4. Full-text search (backend + frontend)
+5. Webhook system for room events
+6. Environment variable migration
+7. Security fixes and auth improvements
+8. Docker production frontend
+9. Meeting user ID removal (schema change)
+10. NextJS version upgrades
+
+**High conflict risk files:**
+- `server/reflector/views/rooms.py` - 12x more changes in main
+- `server/reflector/db/rooms.py` - Main added 7+ fields
+- `www/package.json` - NextJS major version bump
+- Database migrations - 20+ new migrations in main
+
+## Implementation Approach
+
+### Phase 1: Copy Clean Abstractions (1-2 hours)
+
+**Files to copy directly from reference:**
+```bash
+# Core abstraction (review but mostly safe to copy)
+cp -r reflector-dailyco-reference/server/reflector/video_platforms/ \
+      server/reflector/
+
+# BUT review each file for:
+# - Import paths (make sure they match current main)
+# - Settings references (adapt to current settings.py)
+# - Type imports (ensure no circular dependencies)
+```
+
+**After copying, immediately:**
+```bash
+cd server
+# Check for issues
+uv run ruff check reflector/video_platforms/
+# Fix any import errors or type issues
+```
+
+### Phase 2: Adapt to Current Main (2-3 hours)
+
+**2.1 Settings Integration**
+
+File: `server/reflector/settings.py`
+
+Add at the appropriate location (near existing Whereby settings):
+
+```python
+# Daily.co API Integration (NEW)
+DAILY_API_KEY: str | None = None
+DAILY_WEBHOOK_SECRET: str | None = None
+DAILY_SUBDOMAIN: str | None = None
+AWS_DAILY_S3_BUCKET: str | None = None
+AWS_DAILY_S3_REGION: str = "us-west-2"
+AWS_DAILY_ROLE_ARN: str | None = None
+
+# Platform Migration Feature Flags (NEW)
+DAILY_MIGRATION_ENABLED: bool = False  # Conservative default
+DAILY_MIGRATION_ROOM_IDS: list[str] = []
+DEFAULT_VIDEO_PLATFORM: Literal["whereby", "daily"] = "whereby"
+```
+
+**2.2 Database Migration**
+
+⚠️ **CRITICAL: Do NOT copy migration from reference**
+
+Generate new migration:
+```bash
+cd server
+uv run alembic revision -m "add_platform_support"
+```
+
+Edit the generated migration file to add `platform` column:
+```python
+def upgrade():
+    with op.batch_alter_table("room", schema=None) as batch_op:
+        batch_op.add_column(
+            sa.Column("platform", sa.String(), nullable=False, server_default="whereby")
+        )
+
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.add_column(
+            sa.Column("platform", sa.String(), nullable=False, server_default="whereby")
+        )
+```
+
+**2.3 Update Database Models**
+
+File: `server/reflector/db/rooms.py`
+
+Add platform field (preserve all existing fields from main):
+```python
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from reflector.video_platforms.models import Platform
+
+class Room:
+    # ... ALL existing fields from main (calendar, webhooks, etc.) ...
+
+    # NEW: Platform field
+    platform: "Platform" = sqlalchemy.Column(
+        sqlalchemy.String,
+        nullable=False,
+        server_default="whereby",
+    )
+```
+
+File: `server/reflector/db/meetings.py`
+
+Same approach - add platform field, preserve everything from main.
+
+**2.4 Integrate Platform Abstraction into rooms.py**
+
+⚠️ **This is the most delicate part - main has calendar/webhook features**
+
+File: `server/reflector/views/rooms.py`
+
+Strategy:
+1. Add imports at top
+2. Modify meeting creation logic only
+3. Preserve all calendar/webhook/ICS logic from main
+
+```python
+# Add imports
+from reflector.video_platforms import (
+    create_platform_client,
+    get_platform_for_room,
+)
+
+# In create_meeting endpoint:
+# OLD: Direct Whereby API calls
+# NEW: Platform abstraction
+
+# Find the meeting creation section and replace:
+platform = get_platform_for_room(room.id)
+client = create_platform_client(platform)
+
+meeting_data = await client.create_meeting(
+    room_name_prefix=room.name,
+    end_date=meeting_data.end_date,
+    room=room,
+)
+
+# Then create Meeting record with meeting_data.platform, meeting_data.meeting_id, etc.
+```
+
+**2.5 Add Daily.co Webhook Handler**
+
+Copy from reference, minimal changes needed:
+```bash
+cp reflector-dailyco-reference/server/reflector/views/daily.py \
+   server/reflector/views/
+```
+
+Register in `server/reflector/app.py`:
+```python
+from reflector.views import daily
+
+app.include_router(daily.router, prefix="/v1/daily", tags=["daily"])
+```
+
+**2.6 Add Recording Processing Task**
+
+File: `server/reflector/worker/process.py`
+
+Add the `process_recording_from_url` task from reference (copy the function).
+
+### Phase 3: Frontend Adaptation (3-4 hours)
+
+**3.1 Determine Current API Client Pattern**
+
+First, check how main currently makes API calls:
+```bash
+cd www
+grep -r "api\." app/ | head -20
+# Look for patterns like: api.v1Something()
+```
+
+**3.2 Create Components**
+
+Copy component structure from reference but **rewrite all API calls**:
+
+```bash
+mkdir -p www/app/[roomName]/components
+```
+
+Files to create:
+- `RoomContainer.tsx` - Platform router (mostly copy-safe, just fix imports)
+- `DailyRoom.tsx` - Needs React Query API calls
+- `WherebyRoom.tsx` - Extract current room page logic
+
+**Example React Query pattern** (adapt to your actual API):
+```typescript
+import { api } from '@/app/api/client'
+
+// In DailyRoom.tsx
+const handleConsent = async () => {
+  try {
+    await api.v1MeetingAudioConsent({
+      path: { meeting_id: meeting.id },
+      body: { consent: true },
+    })
+    // ...
+  } catch (error) {
+    // ...
+  }
+}
+```
+
+**3.3 Add Daily.co Dependency**
+
+Check current package manager:
+```bash
+cd www
+ls package-lock.json yarn.lock pnpm-lock.yaml
+```
+
+Then install:
+```bash
+# If using pnpm
+pnpm add @daily-co/daily-js@^0.81.0
+
+# If using yarn
+yarn add @daily-co/daily-js@^0.81.0
+```
+
+**3.4 Update TypeScript Types**
+
+After backend changes, regenerate types:
+```bash
+cd www
+pnpm openapi  # or yarn openapi
+```
+
+This should pick up the new `platform` field on Meeting type.
+
+### Phase 4: Testing (2-3 hours)
+
+**4.1 Copy Test Structure**
+
+```bash
+cp reflector-dailyco-reference/server/tests/test_video_platforms.py \
+   server/tests/
+
+cp reflector-dailyco-reference/server/tests/test_daily_webhook.py \
+   server/tests/
+```
+
+**4.2 Fix Test Imports and Fixtures**
+
+Update imports to match current test infrastructure:
+- Check `server/tests/conftest.py` for fixture patterns
+- Update database access patterns if changed
+- Fix any import errors
+
+**4.3 Run Tests**
+
+```bash
+cd server
+# Run with environment variables for Mac
+REDIS_HOST=localhost \
+CELERY_BROKER_URL=redis://localhost:6379/1 \
+CELERY_RESULT_BACKEND=redis://localhost:6379/1 \
+uv run pytest tests/test_video_platforms.py -v
+```
+
+### Phase 5: Environment Configuration
+
+**Update `server/env.example`:**
+
+Add at the end:
+```bash
+# Daily.co API Integration
+DAILY_API_KEY=your-daily-api-key
+DAILY_WEBHOOK_SECRET=your-daily-webhook-secret
+DAILY_SUBDOMAIN=your-subdomain
+AWS_DAILY_S3_BUCKET=your-daily-bucket
+AWS_DAILY_S3_REGION=us-west-2
+AWS_DAILY_ROLE_ARN=arn:aws:iam::ACCOUNT:role/DailyRecording
+
+# Platform Selection
+DAILY_MIGRATION_ENABLED=false           # Master switch
+DAILY_MIGRATION_ROOM_IDS=[]            # Specific room IDs
+DEFAULT_VIDEO_PLATFORM=whereby          # Default platform
+```
+
+## Decision Tree: Copy vs Adapt vs Rewrite
+
+```
+┌─ Is it pure abstraction logic? (base.py, registry.py, models.py)
+│  YES → Copy directly, review imports
+│  NO  → Continue ↓
+│
+├─ Does it touch database models?
+│  YES → Adapt carefully, preserve main's fields
+│  NO  → Continue ↓
+│
+├─ Does it make API calls on frontend?
+│  YES → Rewrite using React Query
+│  NO  → Continue ↓
+│
+├─ Is it a database migration?
+│  YES → Generate fresh from current schema
+│  NO  → Continue ↓
+│
+└─ Does it touch rooms.py or core business logic?
+   YES → Merge carefully, preserve calendar/webhooks
+   NO  → Safe to adapt from reference
+```
+
+## Verification Checklist
+
+After each phase, verify:
+
+**Phase 1 (Abstraction Layer):**
+- [ ] `uv run ruff check server/reflector/video_platforms/` passes
+- [ ] No circular import errors
+- [ ] Can import `from reflector.video_platforms import create_platform_client`
+
+**Phase 2 (Backend Integration):**
+- [ ] `uv run ruff check server/` passes
+- [ ] Migration file generated (not copied)
+- [ ] Room and Meeting models have platform field
+- [ ] rooms.py still has calendar/webhook features
+
+**Phase 3 (Frontend):**
+- [ ] `pnpm lint` passes
+- [ ] No TypeScript errors
+- [ ] No `@ts-ignore` for platform field
+- [ ] API calls use React Query patterns
+
+**Phase 4 (Testing):**
+- [ ] Tests can be collected: `pytest tests/test_video_platforms.py --collect-only`
+- [ ] Database fixtures work
+- [ ] Mock platform works
+
+**Phase 5 (Config):**
+- [ ] env.example has Daily.co variables
+- [ ] settings.py has all new variables
+- [ ] No duplicate variable definitions
+
+## Common Pitfalls
+
+### 1. Database Schema Conflicts
+**Problem:** Reference removed fields that main has (calendar, webhooks)
+**Solution:** Always preserve main's fields, only add platform field
+
+### 2. Migration Conflicts
+**Problem:** Reference migration has wrong `down_revision`
+**Solution:** Always generate fresh migration from current main
+
+### 3. Frontend API Calls
+**Problem:** Reference uses old API client patterns
+**Solution:** Check current main's API usage, replicate that pattern
+
+### 4. Import Errors
+**Problem:** Circular imports with TYPE_CHECKING
+**Solution:** Use `if TYPE_CHECKING:` for Room/Meeting imports in video_platforms
+
+### 5. Test Database Issues
+**Problem:** Tests fail with "could not translate host name 'postgres'"
+**Solution:** Use environment variables: `REDIS_HOST=localhost DATABASE_URL=...`
+
+### 6. Preserved Features Broken
+**Problem:** Calendar/webhook features stop working
+**Solution:** Carefully review rooms.py diff, only change meeting creation, not calendar logic
+
+## File Modification Summary
+
+**New files (can copy):**
+- `server/reflector/video_platforms/*.py` (entire directory)
+- `server/reflector/views/daily.py`
+- `server/tests/test_video_platforms.py`
+- `server/tests/test_daily_webhook.py`
+- `www/app/[roomName]/components/RoomContainer.tsx`
+- `www/app/[roomName]/components/DailyRoom.tsx`
+- `www/app/[roomName]/components/WherebyRoom.tsx`
+
+**Modified files (careful merging):**
+- `server/reflector/settings.py` - Add Daily.co settings
+- `server/reflector/db/rooms.py` - Add platform field
+- `server/reflector/db/meetings.py` - Add platform field
+- `server/reflector/views/rooms.py` - Integrate platform abstraction
+- `server/reflector/worker/process.py` - Add process_recording_from_url
+- `server/reflector/app.py` - Register daily router
+- `server/env.example` - Add Daily.co variables
+- `www/app/[roomName]/page.tsx` - Use RoomContainer
+- `www/package.json` - Add @daily-co/daily-js
+
+**Generated files (do not copy):**
+- `server/migrations/versions/XXXXXX_add_platform_support.py` - Generate fresh
+
+## Success Metrics
+
+Implementation is complete when:
+- [ ] All tests pass (including new platform tests)
+- [ ] Linting passes (ruff, pnpm lint)
+- [ ] Migration applies cleanly: `uv run alembic upgrade head`
+- [ ] Can create Whereby meeting (existing flow unchanged)
+- [ ] Can create Daily.co meeting (with env vars set)
+- [ ] Frontend loads without TypeScript errors
+- [ ] No features from main were accidentally removed
+
+## Getting Help
+
+**Reference documentation locations:**
+- Implementation plan: `PLAN.md`
+- Reference implementation: `./reflector-dailyco-reference/`
+- Current main codebase: `./ ` (current directory)
+
+**Compare implementations:**
+```bash
+# Compare specific files
+diff reflector-dailyco-reference/server/reflector/video_platforms/base.py \
+     server/reflector/video_platforms/base.py
+
+# See what changed in rooms.py between reference branch point and now
+git log --oneline --since="2025-08-01" -- server/reflector/views/rooms.py
+```
+
+**Key insight:** The reference branch validates the approach and provides working code patterns, but you're implementing fresh against current main to avoid merge conflicts and preserve all new features.

README.md

@@ -1,43 +1,60 @@
 <div align="center">
+<img width="100" alt="image" src="https://github.com/user-attachments/assets/66fb367b-2c89-4516-9912-f47ac59c6a7f"/>
 
 # Reflector
 
-Reflector Audio Management and Analysis is a cutting-edge web application under development by Monadical. It utilizes AI to record meetings, providing a permanent record with transcripts, translations, and automated summaries.
+Reflector is an AI-powered audio transcription and meeting analysis platform that provides real-time transcription, speaker diarization, translation and summarization for audio content and live meetings. It works 100% with local models (whisper/parakeet, pyannote, seamless-m4t, and your local llm like phi-4).
 
-[![Tests](https://github.com/monadical-sas/reflector/actions/workflows/pytests.yml/badge.svg?branch=main&event=push)](https://github.com/monadical-sas/reflector/actions/workflows/pytests.yml)
+[![Tests](https://github.com/monadical-sas/reflector/actions/workflows/test_server.yml/badge.svg?branch=main&event=push)](https://github.com/monadical-sas/reflector/actions/workflows/test_server.yml)
 [![License: MIT](https://img.shields.io/badge/license-MIT-green.svg)](https://opensource.org/licenses/MIT)
 </div>
-
-## Screenshots
+</div>
 <table>
   <tr>
     <td>
-      <a href="https://github.com/user-attachments/assets/3a976930-56c1-47ef-8c76-55d3864309e3">
-        <img width="700" alt="image" src="https://github.com/user-attachments/assets/3a976930-56c1-47ef-8c76-55d3864309e3" />
+      <a href="https://github.com/user-attachments/assets/21f5597c-2930-4899-a154-f7bd61a59e97">
+        <img width="700" alt="image" src="https://github.com/user-attachments/assets/21f5597c-2930-4899-a154-f7bd61a59e97" />
       </a>
     </td>
     <td>
-      <a href="https://github.com/user-attachments/assets/bfe3bde3-08af-4426-a9a1-11ad5cd63b33">
-        <img width="700" alt="image" src="https://github.com/user-attachments/assets/bfe3bde3-08af-4426-a9a1-11ad5cd63b33" />
+      <a href="https://github.com/user-attachments/assets/f6b9399a-5e51-4bae-b807-59128d0a940c">
+        <img width="700" alt="image" src="https://github.com/user-attachments/assets/f6b9399a-5e51-4bae-b807-59128d0a940c" />
       </a>
     </td>
     <td>
-      <a href="https://github.com/user-attachments/assets/7b60c9d0-efe4-474f-a27b-ea13bd0fabdc">
-        <img width="700" alt="image" src="https://github.com/user-attachments/assets/7b60c9d0-efe4-474f-a27b-ea13bd0fabdc" />
+      <a href="https://github.com/user-attachments/assets/a42ce460-c1fd-4489-a995-270516193897">
+        <img width="700" alt="image" src="https://github.com/user-attachments/assets/a42ce460-c1fd-4489-a995-270516193897" />
+      </a>
+    </td>
+    <td>
+      <a href="https://github.com/user-attachments/assets/21929f6d-c309-42fe-9c11-f1299e50fbd4">
+        <img width="700" alt="image" src="https://github.com/user-attachments/assets/21929f6d-c309-42fe-9c11-f1299e50fbd4" />
       </a>
     </td>
   </tr>
 </table>
 
+## What is Reflector?
+
+Reflector is a web application that utilizes local models to process audio content, providing:
+
+- **Real-time Transcription**: Convert speech to text using [Whisper](https://github.com/openai/whisper) (multi-language) or [Parakeet](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2) (English) models
+- **Speaker Diarization**: Identify and label different speakers using [Pyannote](https://github.com/pyannote/pyannote-audio) 3.1
+- **Live Translation**: Translate audio content in real-time to many languages with [Facebook Seamless-M4T](https://github.com/facebookresearch/seamless_communication)
+- **Topic Detection & Summarization**: Extract key topics and generate concise summaries using LLMs
+- **Meeting Recording**: Create permanent records of meetings with searchable transcripts
+
+Currently we provide [modal.com](https://modal.com/) gpu template to deploy.
+
 ## Background
 
 The project architecture consists of three primary components:
 
-- **Front-End**: NextJS React project hosted on Vercel, located in `www/`.
 - **Back-End**: Python server that offers an API and data persistence, found in `server/`.
-- **GPU implementation**: Providing services such as speech-to-text transcription, topic generation, automated summaries, and translations. Most reliable option is Modal deployment
+- **Front-End**: NextJS React project hosted on Vercel, located in `www/`.
+- **GPU implementation**: Providing services such as speech-to-text transcription, topic generation, automated summaries, and translations.
 
-It also uses authentik for authentication if activated, and Vercel for deployment and configuration of the front-end.
+It also uses authentik for authentication if activated.
 
 ## Contribution Guidelines
 
@@ -72,6 +89,8 @@ Note: We currently do not have instructions for Windows users.
 
 ## Installation
 
+*Note: we're working toward better installation, theses instructions are not accurate for now*
+
 ### Frontend
 
 Start with `cd www`.
@@ -80,11 +99,10 @@ Start with `cd www`.
 
 ```bash
 pnpm install
-cp .env_template .env
-cp config-template.ts config.ts
+cp .env.example .env
 ```
 
-Then, fill in the environment variables in `.env` and the configuration in `config.ts` as needed. If you are unsure on how to proceed, ask in Zulip.
+Then, fill in the environment variables in `.env` as needed. If you are unsure on how to proceed, ask in Zulip.
 
 **Run in development mode**
 
@@ -149,3 +167,41 @@ You can manually process an audio file by calling the process tool:
 ```bash
 uv run python -m reflector.tools.process path/to/audio.wav
 ```
+
+## Build-time env variables
+
+Next.js projects are more used to NEXT_PUBLIC_ prefixed buildtime vars. We don't have those for the reason we need to serve a ccustomizable prebuild docker container.
+
+Instead, all the variables are runtime. Variables needed to the frontend are served to the frontend app at initial render.
+
+It also means there's no static prebuild and no static files to serve for js/html.
+
+## Feature Flags
+
+Reflector uses environment variable-based feature flags to control application functionality. These flags allow you to enable or disable features without code changes.
+
+### Available Feature Flags
+
+| Feature Flag | Environment Variable |
+|-------------|---------------------|
+| `requireLogin` | `FEATURE_REQUIRE_LOGIN` |
+| `privacy` | `FEATURE_PRIVACY` |
+| `browse` | `FEATURE_BROWSE` |
+| `sendToZulip` | `FEATURE_SEND_TO_ZULIP` |
+| `rooms` | `FEATURE_ROOMS` |
+
+### Setting Feature Flags
+
+Feature flags are controlled via environment variables using the pattern `FEATURE_{FEATURE_NAME}` where `{FEATURE_NAME}` is the SCREAMING_SNAKE_CASE version of the feature name.
+
+**Examples:**
+```bash
+# Enable user authentication requirement
+FEATURE_REQUIRE_LOGIN=true
+
+# Disable browse functionality
+FEATURE_BROWSE=false
+
+# Enable Zulip integration
+FEATURE_SEND_TO_ZULIP=true
+```

docker-compose.prod.yml

@@ -0,0 +1,39 @@
+# Production Docker Compose configuration for Frontend
+# Usage: docker compose -f docker-compose.prod.yml up -d
+
+services:
+  web:
+    build:
+      context: ./www
+      dockerfile: Dockerfile
+    image: reflector-frontend:latest
+    environment:
+      - KV_URL=${KV_URL:-redis://redis:6379}
+      - SITE_URL=${SITE_URL}
+      - API_URL=${API_URL}
+      - WEBSOCKET_URL=${WEBSOCKET_URL}
+      - NEXTAUTH_URL=${NEXTAUTH_URL:-http://localhost:3000}
+      - NEXTAUTH_SECRET=${NEXTAUTH_SECRET:-changeme-in-production}
+      - AUTHENTIK_ISSUER=${AUTHENTIK_ISSUER}
+      - AUTHENTIK_CLIENT_ID=${AUTHENTIK_CLIENT_ID}
+      - AUTHENTIK_CLIENT_SECRET=${AUTHENTIK_CLIENT_SECRET}
+      - AUTHENTIK_REFRESH_TOKEN_URL=${AUTHENTIK_REFRESH_TOKEN_URL}
+      - SENTRY_DSN=${SENTRY_DSN}
+      - SENTRY_IGNORE_API_RESOLUTION_ERROR=${SENTRY_IGNORE_API_RESOLUTION_ERROR:-1}
+    depends_on:
+      - redis
+    restart: unless-stopped
+
+  redis:
+    image: redis:7.2-alpine
+    restart: unless-stopped
+    healthcheck:
+      test: ["CMD", "redis-cli", "ping"]
+      interval: 30s
+      timeout: 3s
+      retries: 3
+    volumes:
+      - redis_data:/data
+
+volumes:
+  redis_data:

docker-compose.yml

@@ -6,6 +6,7 @@ services:
       - 1250:1250
     volumes:
       - ./server/:/app/
+      - /app/.venv
     env_file:
       - ./server/.env
     environment:
@@ -16,6 +17,7 @@ services:
       context: server
     volumes:
       - ./server/:/app/
+      - /app/.venv
     env_file:
       - ./server/.env
     environment:
@@ -26,6 +28,7 @@ services:
       context: server
     volumes:
       - ./server/:/app/
+      - /app/.venv
     env_file:
       - ./server/.env
     environment:
@@ -36,7 +39,7 @@ services:
     ports:
       - 6379:6379
   web:
-    image: node:18
+    image: node:22-alpine
     ports:
       - "3000:3000"
     command: sh -c "corepack enable && pnpm install && pnpm dev"
@@ -47,6 +50,8 @@ services:
       - /app/node_modules
     env_file:
       - ./www/.env.local
+    environment:
+      - NODE_ENV=development
 
   postgres:
     image: postgres:17

gpu/modal_deployments/.gitignore

@@ -0,0 +1,33 @@
+# OS / Editor
+.DS_Store
+.vscode/
+.idea/
+
+# Python
+__pycache__/
+*.py[cod]
+*$py.class
+
+# Logs
+*.log
+
+# Env and secrets
+.env
+.env.*
+*.env
+*.secret
+
+# Build / dist
+build/
+dist/
+.eggs/
+*.egg-info/
+
+# Coverage / test
+.pytest_cache/
+.coverage*
+htmlcov/
+
+# Modal local state (if any)
+modal_mounts/
+.modal_cache/

gpu/modal_deployments/reflector_transcriber.py

@@ -0,0 +1,608 @@
+import os
+import sys
+import threading
+import uuid
+from typing import Generator, Mapping, NamedTuple, NewType, TypedDict
+from urllib.parse import urlparse
+
+import modal
+
+MODEL_NAME = "large-v2"
+MODEL_COMPUTE_TYPE: str = "float16"
+MODEL_NUM_WORKERS: int = 1
+MINUTES = 60  # seconds
+SAMPLERATE = 16000
+UPLOADS_PATH = "/uploads"
+CACHE_PATH = "/models"
+SUPPORTED_FILE_EXTENSIONS = ["mp3", "mp4", "mpeg", "mpga", "m4a", "wav", "webm"]
+VAD_CONFIG = {
+    "batch_max_duration": 30.0,
+    "silence_padding": 0.5,
+    "window_size": 512,
+}
+
+
+WhisperUniqFilename = NewType("WhisperUniqFilename", str)
+AudioFileExtension = NewType("AudioFileExtension", str)
+
+app = modal.App("reflector-transcriber")
+
+model_cache = modal.Volume.from_name("models", create_if_missing=True)
+upload_volume = modal.Volume.from_name("whisper-uploads", create_if_missing=True)
+
+
+class TimeSegment(NamedTuple):
+    """Represents a time segment with start and end times."""
+
+    start: float
+    end: float
+
+
+class AudioSegment(NamedTuple):
+    """Represents an audio segment with timing and audio data."""
+
+    start: float
+    end: float
+    audio: any
+
+
+class TranscriptResult(NamedTuple):
+    """Represents a transcription result with text and word timings."""
+
+    text: str
+    words: list["WordTiming"]
+
+
+class WordTiming(TypedDict):
+    """Represents a word with its timing information."""
+
+    word: str
+    start: float
+    end: float
+
+
+def download_model():
+    from faster_whisper import download_model
+
+    model_cache.reload()
+
+    download_model(MODEL_NAME, cache_dir=CACHE_PATH)
+
+    model_cache.commit()
+
+
+image = (
+    modal.Image.debian_slim(python_version="3.12")
+    .env(
+        {
+            "HF_HUB_ENABLE_HF_TRANSFER": "1",
+            "LD_LIBRARY_PATH": (
+                "/usr/local/lib/python3.12/site-packages/nvidia/cudnn/lib/:"
+                "/opt/conda/lib/python3.12/site-packages/nvidia/cublas/lib/"
+            ),
+        }
+    )
+    .apt_install("ffmpeg")
+    .pip_install(
+        "huggingface_hub==0.27.1",
+        "hf-transfer==0.1.9",
+        "torch==2.5.1",
+        "faster-whisper==1.1.1",
+        "fastapi==0.115.12",
+        "requests",
+        "librosa==0.10.1",
+        "numpy<2",
+        "silero-vad==5.1.0",
+    )
+    .run_function(download_model, volumes={CACHE_PATH: model_cache})
+)
+
+
+def detect_audio_format(url: str, headers: Mapping[str, str]) -> AudioFileExtension:
+    parsed_url = urlparse(url)
+    url_path = parsed_url.path
+
+    for ext in SUPPORTED_FILE_EXTENSIONS:
+        if url_path.lower().endswith(f".{ext}"):
+            return AudioFileExtension(ext)
+
+    content_type = headers.get("content-type", "").lower()
+    if "audio/mpeg" in content_type or "audio/mp3" in content_type:
+        return AudioFileExtension("mp3")
+    if "audio/wav" in content_type:
+        return AudioFileExtension("wav")
+    if "audio/mp4" in content_type:
+        return AudioFileExtension("mp4")
+
+    raise ValueError(
+        f"Unsupported audio format for URL: {url}. "
+        f"Supported extensions: {', '.join(SUPPORTED_FILE_EXTENSIONS)}"
+    )
+
+
+def download_audio_to_volume(
+    audio_file_url: str,
+) -> tuple[WhisperUniqFilename, AudioFileExtension]:
+    import requests
+    from fastapi import HTTPException
+
+    response = requests.head(audio_file_url, allow_redirects=True)
+    if response.status_code == 404:
+        raise HTTPException(status_code=404, detail="Audio file not found")
+
+    response = requests.get(audio_file_url, allow_redirects=True)
+    response.raise_for_status()
+
+    audio_suffix = detect_audio_format(audio_file_url, response.headers)
+    unique_filename = WhisperUniqFilename(f"{uuid.uuid4()}.{audio_suffix}")
+    file_path = f"{UPLOADS_PATH}/{unique_filename}"
+
+    with open(file_path, "wb") as f:
+        f.write(response.content)
+
+    upload_volume.commit()
+    return unique_filename, audio_suffix
+
+
+def pad_audio(audio_array, sample_rate: int = SAMPLERATE):
+    """Add 0.5s of silence if audio is shorter than the silence_padding window.
+
+    Whisper does not require this strictly, but aligning behavior with Parakeet
+    avoids edge-case crashes on extremely short inputs and makes comparisons easier.
+    """
+    import numpy as np
+
+    audio_duration = len(audio_array) / sample_rate
+    if audio_duration < VAD_CONFIG["silence_padding"]:
+        silence_samples = int(sample_rate * VAD_CONFIG["silence_padding"])
+        silence = np.zeros(silence_samples, dtype=np.float32)
+        return np.concatenate([audio_array, silence])
+    return audio_array
+
+
+@app.cls(
+    gpu="A10G",
+    timeout=5 * MINUTES,
+    scaledown_window=5 * MINUTES,
+    image=image,
+    volumes={CACHE_PATH: model_cache, UPLOADS_PATH: upload_volume},
+)
+@modal.concurrent(max_inputs=10)
+class TranscriberWhisperLive:
+    """Live transcriber class for small audio segments (A10G).
+
+    Mirrors the Parakeet live class API but uses Faster-Whisper under the hood.
+    """
+
+    @modal.enter()
+    def enter(self):
+        import faster_whisper
+        import torch
+
+        self.lock = threading.Lock()
+        self.use_gpu = torch.cuda.is_available()
+        self.device = "cuda" if self.use_gpu else "cpu"
+        self.model = faster_whisper.WhisperModel(
+            MODEL_NAME,
+            device=self.device,
+            compute_type=MODEL_COMPUTE_TYPE,
+            num_workers=MODEL_NUM_WORKERS,
+            download_root=CACHE_PATH,
+            local_files_only=True,
+        )
+        print(f"Model is on device: {self.device}")
+
+    @modal.method()
+    def transcribe_segment(
+        self,
+        filename: str,
+        language: str = "en",
+    ):
+        """Transcribe a single uploaded audio file by filename."""
+        upload_volume.reload()
+
+        file_path = f"{UPLOADS_PATH}/{filename}"
+        if not os.path.exists(file_path):
+            raise FileNotFoundError(f"File not found: {file_path}")
+
+        with self.lock:
+            with NoStdStreams():
+                segments, _ = self.model.transcribe(
+                    file_path,
+                    language=language,
+                    beam_size=5,
+                    word_timestamps=True,
+                    vad_filter=True,
+                    vad_parameters={"min_silence_duration_ms": 500},
+                )
+
+        segments = list(segments)
+        text = "".join(segment.text for segment in segments).strip()
+        words = [
+            {
+                "word": word.word,
+                "start": round(float(word.start), 2),
+                "end": round(float(word.end), 2),
+            }
+            for segment in segments
+            for word in segment.words
+        ]
+
+        return {"text": text, "words": words}
+
+    @modal.method()
+    def transcribe_batch(
+        self,
+        filenames: list[str],
+        language: str = "en",
+    ):
+        """Transcribe multiple uploaded audio files and return per-file results."""
+        upload_volume.reload()
+
+        results = []
+        for filename in filenames:
+            file_path = f"{UPLOADS_PATH}/{filename}"
+            if not os.path.exists(file_path):
+                raise FileNotFoundError(f"Batch file not found: {file_path}")
+
+            with self.lock:
+                with NoStdStreams():
+                    segments, _ = self.model.transcribe(
+                        file_path,
+                        language=language,
+                        beam_size=5,
+                        word_timestamps=True,
+                        vad_filter=True,
+                        vad_parameters={"min_silence_duration_ms": 500},
+                    )
+
+            segments = list(segments)
+            text = "".join(seg.text for seg in segments).strip()
+            words = [
+                {
+                    "word": w.word,
+                    "start": round(float(w.start), 2),
+                    "end": round(float(w.end), 2),
+                }
+                for seg in segments
+                for w in seg.words
+            ]
+
+            results.append(
+                {
+                    "filename": filename,
+                    "text": text,
+                    "words": words,
+                }
+            )
+
+        return results
+
+
+@app.cls(
+    gpu="L40S",
+    timeout=15 * MINUTES,
+    image=image,
+    volumes={CACHE_PATH: model_cache, UPLOADS_PATH: upload_volume},
+)
+class TranscriberWhisperFile:
+    """File transcriber for larger/longer audio, using VAD-driven batching (L40S)."""
+
+    @modal.enter()
+    def enter(self):
+        import faster_whisper
+        import torch
+        from silero_vad import load_silero_vad
+
+        self.lock = threading.Lock()
+        self.use_gpu = torch.cuda.is_available()
+        self.device = "cuda" if self.use_gpu else "cpu"
+        self.model = faster_whisper.WhisperModel(
+            MODEL_NAME,
+            device=self.device,
+            compute_type=MODEL_COMPUTE_TYPE,
+            num_workers=MODEL_NUM_WORKERS,
+            download_root=CACHE_PATH,
+            local_files_only=True,
+        )
+        self.vad_model = load_silero_vad(onnx=False)
+
+    @modal.method()
+    def transcribe_segment(
+        self, filename: str, timestamp_offset: float = 0.0, language: str = "en"
+    ):
+        import librosa
+        import numpy as np
+        from silero_vad import VADIterator
+
+        def vad_segments(
+            audio_array,
+            sample_rate: int = SAMPLERATE,
+            window_size: int = VAD_CONFIG["window_size"],
+        ) -> Generator[TimeSegment, None, None]:
+            """Generate speech segments as TimeSegment using Silero VAD."""
+            iterator = VADIterator(self.vad_model, sampling_rate=sample_rate)
+            start = None
+            for i in range(0, len(audio_array), window_size):
+                chunk = audio_array[i : i + window_size]
+                if len(chunk) < window_size:
+                    chunk = np.pad(
+                        chunk, (0, window_size - len(chunk)), mode="constant"
+                    )
+                speech = iterator(chunk)
+                if not speech:
+                    continue
+                if "start" in speech:
+                    start = speech["start"]
+                    continue
+                if "end" in speech and start is not None:
+                    end = speech["end"]
+                    yield TimeSegment(
+                        start / float(SAMPLERATE), end / float(SAMPLERATE)
+                    )
+                    start = None
+            iterator.reset_states()
+
+        upload_volume.reload()
+        file_path = f"{UPLOADS_PATH}/{filename}"
+        if not os.path.exists(file_path):
+            raise FileNotFoundError(f"File not found: {file_path}")
+
+        audio_array, _sr = librosa.load(file_path, sr=SAMPLERATE, mono=True)
+
+        # Batch segments up to ~30s windows by merging contiguous VAD segments
+        merged_batches: list[TimeSegment] = []
+        batch_start = None
+        batch_end = None
+        max_duration = VAD_CONFIG["batch_max_duration"]
+        for segment in vad_segments(audio_array):
+            seg_start, seg_end = segment.start, segment.end
+            if batch_start is None:
+                batch_start, batch_end = seg_start, seg_end
+                continue
+            if seg_end - batch_start <= max_duration:
+                batch_end = seg_end
+            else:
+                merged_batches.append(TimeSegment(batch_start, batch_end))
+                batch_start, batch_end = seg_start, seg_end
+        if batch_start is not None and batch_end is not None:
+            merged_batches.append(TimeSegment(batch_start, batch_end))
+
+        all_text = []
+        all_words = []
+
+        for segment in merged_batches:
+            start_time, end_time = segment.start, segment.end
+            s_idx = int(start_time * SAMPLERATE)
+            e_idx = int(end_time * SAMPLERATE)
+            segment = audio_array[s_idx:e_idx]
+            segment = pad_audio(segment, SAMPLERATE)
+
+            with self.lock:
+                segments, _ = self.model.transcribe(
+                    segment,
+                    language=language,
+                    beam_size=5,
+                    word_timestamps=True,
+                    vad_filter=True,
+                    vad_parameters={"min_silence_duration_ms": 500},
+                )
+
+            segments = list(segments)
+            text = "".join(seg.text for seg in segments).strip()
+            words = [
+                {
+                    "word": w.word,
+                    "start": round(float(w.start) + start_time + timestamp_offset, 2),
+                    "end": round(float(w.end) + start_time + timestamp_offset, 2),
+                }
+                for seg in segments
+                for w in seg.words
+            ]
+            if text:
+                all_text.append(text)
+            all_words.extend(words)
+
+        return {"text": " ".join(all_text), "words": all_words}
+
+
+def detect_audio_format(url: str, headers: dict) -> str:
+    from urllib.parse import urlparse
+
+    from fastapi import HTTPException
+
+    url_path = urlparse(url).path
+    for ext in SUPPORTED_FILE_EXTENSIONS:
+        if url_path.lower().endswith(f".{ext}"):
+            return ext
+
+    content_type = headers.get("content-type", "").lower()
+    if "audio/mpeg" in content_type or "audio/mp3" in content_type:
+        return "mp3"
+    if "audio/wav" in content_type:
+        return "wav"
+    if "audio/mp4" in content_type:
+        return "mp4"
+
+    raise HTTPException(
+        status_code=400,
+        detail=(
+            f"Unsupported audio format for URL. Supported extensions: {', '.join(SUPPORTED_FILE_EXTENSIONS)}"
+        ),
+    )
+
+
+def download_audio_to_volume(audio_file_url: str) -> tuple[str, str]:
+    import requests
+    from fastapi import HTTPException
+
+    response = requests.head(audio_file_url, allow_redirects=True)
+    if response.status_code == 404:
+        raise HTTPException(status_code=404, detail="Audio file not found")
+
+    response = requests.get(audio_file_url, allow_redirects=True)
+    response.raise_for_status()
+
+    audio_suffix = detect_audio_format(audio_file_url, response.headers)
+    unique_filename = f"{uuid.uuid4()}.{audio_suffix}"
+    file_path = f"{UPLOADS_PATH}/{unique_filename}"
+
+    with open(file_path, "wb") as f:
+        f.write(response.content)
+
+    upload_volume.commit()
+    return unique_filename, audio_suffix
+
+
+@app.function(
+    scaledown_window=60,
+    timeout=600,
+    secrets=[
+        modal.Secret.from_name("reflector-gpu"),
+    ],
+    volumes={CACHE_PATH: model_cache, UPLOADS_PATH: upload_volume},
+    image=image,
+)
+@modal.concurrent(max_inputs=40)
+@modal.asgi_app()
+def web():
+    from fastapi import (
+        Body,
+        Depends,
+        FastAPI,
+        Form,
+        HTTPException,
+        UploadFile,
+        status,
+    )
+    from fastapi.security import OAuth2PasswordBearer
+
+    transcriber_live = TranscriberWhisperLive()
+    transcriber_file = TranscriberWhisperFile()
+
+    app = FastAPI()
+
+    oauth2_scheme = OAuth2PasswordBearer(tokenUrl="token")
+
+    def apikey_auth(apikey: str = Depends(oauth2_scheme)):
+        if apikey == os.environ["REFLECTOR_GPU_APIKEY"]:
+            return
+        raise HTTPException(
+            status_code=status.HTTP_401_UNAUTHORIZED,
+            detail="Invalid API key",
+            headers={"WWW-Authenticate": "Bearer"},
+        )
+
+    class TranscriptResponse(dict):
+        pass
+
+    @app.post("/v1/audio/transcriptions", dependencies=[Depends(apikey_auth)])
+    def transcribe(
+        file: UploadFile = None,
+        files: list[UploadFile] | None = None,
+        model: str = Form(MODEL_NAME),
+        language: str = Form("en"),
+        batch: bool = Form(False),
+    ):
+        if not file and not files:
+            raise HTTPException(
+                status_code=400, detail="Either 'file' or 'files' parameter is required"
+            )
+        if batch and not files:
+            raise HTTPException(
+                status_code=400, detail="Batch transcription requires 'files'"
+            )
+
+        upload_files = [file] if file else files
+
+        uploaded_filenames: list[str] = []
+        for upload_file in upload_files:
+            audio_suffix = upload_file.filename.split(".")[-1]
+            if audio_suffix not in SUPPORTED_FILE_EXTENSIONS:
+                raise HTTPException(
+                    status_code=400,
+                    detail=(
+                        f"Unsupported audio format. Supported extensions: {', '.join(SUPPORTED_FILE_EXTENSIONS)}"
+                    ),
+                )
+
+            unique_filename = f"{uuid.uuid4()}.{audio_suffix}"
+            file_path = f"{UPLOADS_PATH}/{unique_filename}"
+            with open(file_path, "wb") as f:
+                content = upload_file.file.read()
+                f.write(content)
+            uploaded_filenames.append(unique_filename)
+
+        upload_volume.commit()
+
+        try:
+            if batch and len(upload_files) > 1:
+                func = transcriber_live.transcribe_batch.spawn(
+                    filenames=uploaded_filenames,
+                    language=language,
+                )
+                results = func.get()
+                return {"results": results}
+
+            results = []
+            for filename in uploaded_filenames:
+                func = transcriber_live.transcribe_segment.spawn(
+                    filename=filename,
+                    language=language,
+                )
+                result = func.get()
+                result["filename"] = filename
+                results.append(result)
+
+            return {"results": results} if len(results) > 1 else results[0]
+        finally:
+            for filename in uploaded_filenames:
+                try:
+                    file_path = f"{UPLOADS_PATH}/{filename}"
+                    os.remove(file_path)
+                except Exception:
+                    pass
+            upload_volume.commit()
+
+    @app.post("/v1/audio/transcriptions-from-url", dependencies=[Depends(apikey_auth)])
+    def transcribe_from_url(
+        audio_file_url: str = Body(
+            ..., description="URL of the audio file to transcribe"
+        ),
+        model: str = Body(MODEL_NAME),
+        language: str = Body("en"),
+        timestamp_offset: float = Body(0.0),
+    ):
+        unique_filename, _audio_suffix = download_audio_to_volume(audio_file_url)
+        try:
+            func = transcriber_file.transcribe_segment.spawn(
+                filename=unique_filename,
+                timestamp_offset=timestamp_offset,
+                language=language,
+            )
+            result = func.get()
+            return result
+        finally:
+            try:
+                file_path = f"{UPLOADS_PATH}/{unique_filename}"
+                os.remove(file_path)
+                upload_volume.commit()
+            except Exception:
+                pass
+
+    return app
+
+
+class NoStdStreams:
+    def __init__(self):
+        self.devnull = open(os.devnull, "w")
+
+    def __enter__(self):
+        self._stdout, self._stderr = sys.stdout, sys.stderr
+        self._stdout.flush()
+        self._stderr.flush()
+        sys.stdout, sys.stderr = self.devnull, self.devnull
+
+    def __exit__(self, exc_type, exc_value, traceback):
+        sys.stdout, sys.stderr = self._stdout, self._stderr
+        self.devnull.close()

gpu/modal_deployments/reflector_transcriber_parakeet.py

@@ -3,7 +3,7 @@ import os
 import sys
 import threading
 import uuid
-from typing import Mapping, NewType
+from typing import Generator, Mapping, NamedTuple, NewType, TypedDict
 from urllib.parse import urlparse
 
 import modal
@@ -14,10 +14,7 @@ SAMPLERATE = 16000
 UPLOADS_PATH = "/uploads"
 CACHE_PATH = "/cache"
 VAD_CONFIG = {
-    "max_segment_duration": 30.0,
-    "batch_max_files": 10,
-    "batch_max_duration": 5.0,
-    "min_segment_duration": 0.02,
+    "batch_max_duration": 30.0,
     "silence_padding": 0.5,
     "window_size": 512,
 }
@@ -25,6 +22,37 @@ VAD_CONFIG = {
 ParakeetUniqFilename = NewType("ParakeetUniqFilename", str)
 AudioFileExtension = NewType("AudioFileExtension", str)
 
+
+class TimeSegment(NamedTuple):
+    """Represents a time segment with start and end times."""
+
+    start: float
+    end: float
+
+
+class AudioSegment(NamedTuple):
+    """Represents an audio segment with timing and audio data."""
+
+    start: float
+    end: float
+    audio: any
+
+
+class TranscriptResult(NamedTuple):
+    """Represents a transcription result with text and word timings."""
+
+    text: str
+    words: list["WordTiming"]
+
+
+class WordTiming(TypedDict):
+    """Represents a word with its timing information."""
+
+    word: str
+    start: float
+    end: float
+
+
 app = modal.App("reflector-transcriber-parakeet")
 
 # Volume for caching model weights
@@ -49,7 +77,7 @@ image = (
     .pip_install(
         "hf_transfer==0.1.9",
         "huggingface_hub[hf-xet]==0.31.2",
-        "nemo_toolkit[asr]==2.3.0",
+        "nemo_toolkit[asr]==2.5.0",
         "cuda-python==12.8.0",
         "fastapi==0.115.12",
         "numpy<2",
@@ -170,12 +198,14 @@ class TranscriberParakeetLive:
                 (output,) = self.model.transcribe([padded_audio], timestamps=True)
 
         text = output.text.strip()
-        words = [
-            {
-                "word": word_info["word"],
-                "start": round(word_info["start"], 2),
-                "end": round(word_info["end"], 2),
-            }
+        words: list[WordTiming] = [
+            WordTiming(
+                # XXX the space added here is to match the output of whisper
+                # whisper add space to each words, while parakeet don't
+                word=word_info["word"] + " ",
+                start=round(word_info["start"], 2),
+                end=round(word_info["end"], 2),
+            )
             for word_info in output.timestamp["word"]
         ]
 
@@ -211,12 +241,12 @@ class TranscriberParakeetLive:
         for i, (filename, output) in enumerate(zip(filenames, outputs)):
             text = output.text.strip()
 
-            words = [
-                {
-                    "word": word_info["word"],
-                    "start": round(word_info["start"], 2),
-                    "end": round(word_info["end"], 2),
-                }
+            words: list[WordTiming] = [
+                WordTiming(
+                    word=word_info["word"] + " ",
+                    start=round(word_info["start"], 2),
+                    end=round(word_info["end"], 2),
+                )
                 for word_info in output.timestamp["word"]
             ]
 
@@ -271,7 +301,9 @@ class TranscriberParakeetFile:
             audio_array, sample_rate = librosa.load(file_path, sr=SAMPLERATE, mono=True)
             return audio_array
 
-        def vad_segment_generator(audio_array):
+        def vad_segment_generator(
+            audio_array,
+        ) -> Generator[TimeSegment, None, None]:
             """Generate speech segments using VAD with start/end sample indices"""
             vad_iterator = VADIterator(self.vad_model, sampling_rate=SAMPLERATE)
             window_size = VAD_CONFIG["window_size"]
@@ -297,107 +329,121 @@ class TranscriberParakeetFile:
                     start_time = start / float(SAMPLERATE)
                     end_time = end / float(SAMPLERATE)
 
-                    # Extract the actual audio segment
-                    audio_segment = audio_array[start:end]
-
-                    yield (start_time, end_time, audio_segment)
+                    yield TimeSegment(start_time, end_time)
                     start = None
 
             vad_iterator.reset_states()
 
-        def vad_segment_filter(segments):
-            """Filter VAD segments by duration and chunk large segments"""
-            min_dur = VAD_CONFIG["min_segment_duration"]
-            max_dur = VAD_CONFIG["max_segment_duration"]
+        def batch_speech_segments(
+            segments: Generator[TimeSegment, None, None], max_duration: int
+        ) -> Generator[TimeSegment, None, None]:
+            """
+            Input segments:
+              [0-2] [3-5] [6-8] [10-11] [12-15] [17-19] [20-22]
 
-            for start_time, end_time, audio_segment in segments:
-                segment_duration = end_time - start_time
+                                  ↓ (max_duration=10)
 
-                # Skip very small segments
-                if segment_duration < min_dur:
+              Output batches:
+              [0-8]           [10-19]          [20-22]
+
+            Note: silences are kept for better transcription, previous implementation was
+            passing segments separatly, but the output was less accurate.
+            """
+            batch_start_time = None
+            batch_end_time = None
+
+            for segment in segments:
+                start_time, end_time = segment.start, segment.end
+                if batch_start_time is None or batch_end_time is None:
+                    batch_start_time = start_time
+                    batch_end_time = end_time
                     continue
 
-                # If segment is within max duration, yield as-is
-                if segment_duration <= max_dur:
-                    yield (start_time, end_time, audio_segment)
+                total_duration = end_time - batch_start_time
+
+                if total_duration <= max_duration:
+                    batch_end_time = end_time
                     continue
 
-                # Chunk large segments into smaller pieces
-                chunk_samples = int(max_dur * SAMPLERATE)
-                current_start = start_time
+                yield TimeSegment(batch_start_time, batch_end_time)
+                batch_start_time = start_time
+                batch_end_time = end_time
 
-                for chunk_offset in range(0, len(audio_segment), chunk_samples):
-                    chunk_audio = audio_segment[
-                        chunk_offset : chunk_offset + chunk_samples
-                    ]
-                    if len(chunk_audio) == 0:
-                        break
+            if batch_start_time is None or batch_end_time is None:
+                return
 
-                    chunk_duration = len(chunk_audio) / float(SAMPLERATE)
-                    chunk_end = current_start + chunk_duration
+            yield TimeSegment(batch_start_time, batch_end_time)
 
-                    # Only yield chunks that meet minimum duration
-                    if chunk_duration >= min_dur:
-                        yield (current_start, chunk_end, chunk_audio)
+        def batch_segment_to_audio_segment(
+            segments: Generator[TimeSegment, None, None],
+            audio_array,
+        ) -> Generator[AudioSegment, None, None]:
+            """Extract audio segments and apply padding for Parakeet compatibility.
 
-                    current_start = chunk_end
+            Uses pad_audio to ensure segments are at least 0.5s long, preventing
+            Parakeet crashes. This padding may cause slight timing overlaps between
+            segments, which are corrected by enforce_word_timing_constraints.
+            """
+            for segment in segments:
+                start_time, end_time = segment.start, segment.end
+                start_sample = int(start_time * SAMPLERATE)
+                end_sample = int(end_time * SAMPLERATE)
+                audio_segment = audio_array[start_sample:end_sample]
 
-        def batch_segments(segments, max_files=10, max_duration=5.0):
-            batch = []
-            batch_duration = 0.0
+                padded_segment = pad_audio(audio_segment, SAMPLERATE)
 
-            for start_time, end_time, audio_segment in segments:
-                segment_duration = end_time - start_time
+                yield AudioSegment(start_time, end_time, padded_segment)
 
-                if segment_duration < VAD_CONFIG["silence_padding"]:
-                    silence_samples = int(
-                        (VAD_CONFIG["silence_padding"] - segment_duration) * SAMPLERATE
-                    )
-                    padding = np.zeros(silence_samples, dtype=np.float32)
-                    audio_segment = np.concatenate([audio_segment, padding])
-                    segment_duration = VAD_CONFIG["silence_padding"]
-
-                batch.append((start_time, end_time, audio_segment))
-                batch_duration += segment_duration
-
-                if len(batch) >= max_files or batch_duration >= max_duration:
-                    yield batch
-                    batch = []
-                    batch_duration = 0.0
-
-            if batch:
-                yield batch
-
-        def transcribe_batch(model, audio_segments):
+        def transcribe_batch(model, audio_segments: list) -> list:
             with NoStdStreams():
                 outputs = model.transcribe(audio_segments, timestamps=True)
             return outputs
 
+        def enforce_word_timing_constraints(
+            words: list[WordTiming],
+        ) -> list[WordTiming]:
+            """Enforce that word end times don't exceed the start time of the next word.
+
+            Due to silence padding added in batch_segment_to_audio_segment for better
+            transcription accuracy, word timings from different segments may overlap.
+            This function ensures there are no overlaps by adjusting end times.
+            """
+            if len(words) <= 1:
+                return words
+
+            enforced_words = []
+            for i, word in enumerate(words):
+                enforced_word = word.copy()
+
+                if i < len(words) - 1:
+                    next_start = words[i + 1]["start"]
+                    if enforced_word["end"] > next_start:
+                        enforced_word["end"] = next_start
+
+                enforced_words.append(enforced_word)
+
+            return enforced_words
+
         def emit_results(
-            results,
-            segments_info,
-            batch_index,
-            total_batches,
-        ):
+            results: list,
+            segments_info: list[AudioSegment],
+        ) -> Generator[TranscriptResult, None, None]:
             """Yield transcribed text and word timings from model output, adjusting timestamps to absolute positions."""
-            for i, (output, (start_time, end_time, _)) in enumerate(
-                zip(results, segments_info)
-            ):
+            for i, (output, segment) in enumerate(zip(results, segments_info)):
+                start_time, end_time = segment.start, segment.end
                 text = output.text.strip()
-                words = [
-                    {
-                        "word": word_info["word"],
-                        "start": round(
+                words: list[WordTiming] = [
+                    WordTiming(
+                        word=word_info["word"] + " ",
+                        start=round(
                             word_info["start"] + start_time + timestamp_offset, 2
                         ),
-                        "end": round(
-                            word_info["end"] + start_time + timestamp_offset, 2
-                        ),
-                    }
+                        end=round(word_info["end"] + start_time + timestamp_offset, 2),
+                    )
                     for word_info in output.timestamp["word"]
                 ]
 
-                yield text, words
+                yield TranscriptResult(text, words)
 
         upload_volume.reload()
 
@@ -407,41 +453,31 @@ class TranscriberParakeetFile:
 
         audio_array = load_and_convert_audio(file_path)
         total_duration = len(audio_array) / float(SAMPLERATE)
-        processed_duration = 0.0
 
-        all_text_parts = []
-        all_words = []
+        all_text_parts: list[str] = []
+        all_words: list[WordTiming] = []
 
         raw_segments = vad_segment_generator(audio_array)
-        filtered_segments = vad_segment_filter(raw_segments)
-        batches = batch_segments(
-            filtered_segments,
-            VAD_CONFIG["batch_max_files"],
+        speech_segments = batch_speech_segments(
+            raw_segments,
             VAD_CONFIG["batch_max_duration"],
         )
+        audio_segments = batch_segment_to_audio_segment(speech_segments, audio_array)
 
-        batch_index = 0
-        total_batches = max(
-            1, int(total_duration / VAD_CONFIG["batch_max_duration"]) + 1
-        )
+        for batch in audio_segments:
+            audio_segment = batch.audio
+            results = transcribe_batch(self.model, [audio_segment])
 
-        for batch in batches:
-            batch_index += 1
-            audio_segments = [seg[2] for seg in batch]
-            results = transcribe_batch(self.model, audio_segments)
-
-            for text, words in emit_results(
+            for result in emit_results(
                 results,
-                batch,
-                batch_index,
-                total_batches,
+                [batch],
             ):
-                if not text:
+                if not result.text:
                     continue
-                all_text_parts.append(text)
-                all_words.extend(words)
+                all_text_parts.append(result.text)
+                all_words.extend(result.words)
 
-            processed_duration += sum(len(seg[2]) / float(SAMPLERATE) for seg in batch)
+        all_words = enforce_word_timing_constraints(all_words)
 
         combined_text = " ".join(all_text_parts)
         return {"text": combined_text, "words": all_words}

gpu/self_hosted/.env.example

@@ -0,0 +1,2 @@
+REFLECTOR_GPU_APIKEY=
+HF_TOKEN=

gpu/self_hosted/.gitignore

@@ -0,0 +1,38 @@
+cache/
+
+# OS / Editor
+.DS_Store
+.vscode/
+.idea/
+
+# Python
+__pycache__/
+*.py[cod]
+*$py.class
+
+# Env and secrets
+.env
+*.env
+*.secret
+HF_TOKEN
+REFLECTOR_GPU_APIKEY
+
+# Virtual env / uv
+.venv/
+venv/
+ENV/
+uv/
+
+# Build / dist
+build/
+dist/
+.eggs/
+*.egg-info/
+
+# Coverage / test
+.pytest_cache/
+.coverage*
+htmlcov/
+
+# Logs
+*.log

gpu/self_hosted/Dockerfile

@@ -0,0 +1,46 @@
+FROM python:3.12-slim
+
+ENV PYTHONUNBUFFERED=1 \
+    UV_LINK_MODE=copy \
+    UV_NO_CACHE=1
+
+WORKDIR /tmp
+RUN apt-get update \
+ && apt-get install -y \
+    ffmpeg \
+    curl \
+    ca-certificates \
+    gnupg \
+    wget \
+ && apt-get clean
+# Add NVIDIA CUDA repo for Debian 12 (bookworm) and install cuDNN 9 for CUDA 12
+ADD https://developer.download.nvidia.com/compute/cuda/repos/debian12/x86_64/cuda-keyring_1.1-1_all.deb /cuda-keyring.deb
+RUN dpkg -i /cuda-keyring.deb \
+ && rm /cuda-keyring.deb \
+ && apt-get update \
+ && apt-get install -y --no-install-recommends \
+    cuda-cudart-12-6 \
+    libcublas-12-6 \
+    libcudnn9-cuda-12 \
+    libcudnn9-dev-cuda-12 \
+ && apt-get clean \
+ && rm -rf /var/lib/apt/lists/*
+ADD https://astral.sh/uv/install.sh /uv-installer.sh
+RUN sh /uv-installer.sh && rm /uv-installer.sh
+ENV PATH="/root/.local/bin/:$PATH"
+ENV LD_LIBRARY_PATH="/usr/local/cuda/lib64:/usr/lib/x86_64-linux-gnu:$LD_LIBRARY_PATH"
+
+RUN mkdir -p /app
+WORKDIR /app
+COPY pyproject.toml uv.lock /app/
+
+
+COPY ./app /app/app
+COPY ./main.py /app/
+COPY ./runserver.sh /app/
+
+EXPOSE 8000
+
+CMD ["sh", "/app/runserver.sh"]
+
+

gpu/self_hosted/README.md

@@ -0,0 +1,73 @@
+# Self-hosted Model API
+
+Run transcription, translation, and diarization services compatible with Reflector's GPU Model API. Works on CPU or GPU.
+
+Environment variables
+
+- REFLECTOR_GPU_APIKEY: Optional Bearer token. If unset, auth is disabled.
+- HF_TOKEN: Optional. Required for diarization to download pyannote pipelines
+
+Requirements
+
+- FFmpeg must be installed and on PATH (used for URL-based and segmented transcription)
+- Python 3.12+
+- NVIDIA GPU optional. If available, it will be used automatically
+
+Local run
+Set env vars in self_hosted/.env file
+uv sync
+
+uv run uvicorn main:app --host 0.0.0.0 --port 8000
+
+Authentication
+
+- If REFLECTOR_GPU_APIKEY is set, include header: Authorization: Bearer <key>
+
+Endpoints
+
+- POST /v1/audio/transcriptions
+
+  - multipart/form-data
+  - fields: file (single file) OR files[] (multiple files), language, batch (true/false)
+  - response: single { text, words, filename } or { results: [ ... ] }
+
+- POST /v1/audio/transcriptions-from-url
+
+  - application/json
+  - body: { audio_file_url, language, timestamp_offset }
+  - response: { text, words }
+
+- POST /translate
+
+  - text: query parameter
+  - body (application/json): { source_language, target_language }
+  - response: { text: { <src>: original, <tgt>: translated } }
+
+- POST /diarize
+  - query parameters: audio_file_url, timestamp (optional)
+  - requires HF_TOKEN to be set (for pyannote)
+  - response: { diarization: [ { start, end, speaker } ] }
+
+OpenAPI docs
+
+- Visit /docs when the server is running
+
+Docker
+
+- Not yet provided in this directory. A Dockerfile will be added later. For now, use Local run above
+
+Conformance tests
+
+# From this directory
+
+TRANSCRIPT_URL=http://localhost:8000 \
+TRANSCRIPT_API_KEY=dev-key \
+uv run -m pytest -m model_api --no-cov ../../server/tests/test_model_api_transcript.py
+
+TRANSLATION_URL=http://localhost:8000 \
+TRANSLATION_API_KEY=dev-key \
+uv run -m pytest -m model_api --no-cov ../../server/tests/test_model_api_translation.py
+
+DIARIZATION_URL=http://localhost:8000 \
+DIARIZATION_API_KEY=dev-key \
+uv run -m pytest -m model_api --no-cov ../../server/tests/test_model_api_diarization.py

gpu/self_hosted/app/auth.py

@@ -0,0 +1,19 @@
+import os
+
+from fastapi import Depends, HTTPException, status
+from fastapi.security import OAuth2PasswordBearer
+
+oauth2_scheme = OAuth2PasswordBearer(tokenUrl="token")
+
+
+def apikey_auth(apikey: str = Depends(oauth2_scheme)):
+    required_key = os.environ.get("REFLECTOR_GPU_APIKEY")
+    if not required_key:
+        return
+    if apikey == required_key:
+        return
+    raise HTTPException(
+        status_code=status.HTTP_401_UNAUTHORIZED,
+        detail="Invalid API key",
+        headers={"WWW-Authenticate": "Bearer"},
+    )

gpu/self_hosted/app/config.py

@@ -0,0 +1,12 @@
+from pathlib import Path
+
+SUPPORTED_FILE_EXTENSIONS = ["mp3", "mp4", "mpeg", "mpga", "m4a", "wav", "webm"]
+SAMPLE_RATE = 16000
+VAD_CONFIG = {
+    "batch_max_duration": 30.0,
+    "silence_padding": 0.5,
+    "window_size": 512,
+}
+
+# App-level paths
+UPLOADS_PATH = Path("/tmp/whisper-uploads")

gpu/self_hosted/app/factory.py

@@ -0,0 +1,30 @@
+from contextlib import asynccontextmanager
+
+from fastapi import FastAPI
+
+from .routers.diarization import router as diarization_router
+from .routers.transcription import router as transcription_router
+from .routers.translation import router as translation_router
+from .services.transcriber import WhisperService
+from .services.diarizer import PyannoteDiarizationService
+from .utils import ensure_dirs
+
+
+@asynccontextmanager
+async def lifespan(app: FastAPI):
+    ensure_dirs()
+    whisper_service = WhisperService()
+    whisper_service.load()
+    app.state.whisper = whisper_service
+    diarization_service = PyannoteDiarizationService()
+    diarization_service.load()
+    app.state.diarizer = diarization_service
+    yield
+
+
+def create_app() -> FastAPI:
+    app = FastAPI(lifespan=lifespan)
+    app.include_router(transcription_router)
+    app.include_router(translation_router)
+    app.include_router(diarization_router)
+    return app

gpu/self_hosted/app/routers/diarization.py

@@ -0,0 +1,30 @@
+from typing import List
+
+from fastapi import APIRouter, Depends, Request
+from pydantic import BaseModel
+
+from ..auth import apikey_auth
+from ..services.diarizer import PyannoteDiarizationService
+from ..utils import download_audio_file
+
+router = APIRouter(tags=["diarization"])
+
+
+class DiarizationSegment(BaseModel):
+    start: float
+    end: float
+    speaker: int
+
+
+class DiarizationResponse(BaseModel):
+    diarization: List[DiarizationSegment]
+
+
+@router.post(
+    "/diarize", dependencies=[Depends(apikey_auth)], response_model=DiarizationResponse
+)
+def diarize(request: Request, audio_file_url: str, timestamp: float = 0.0):
+    with download_audio_file(audio_file_url) as (file_path, _ext):
+        file_path = str(file_path)
+        diarizer: PyannoteDiarizationService = request.app.state.diarizer
+        return diarizer.diarize_file(file_path, timestamp=timestamp)

gpu/self_hosted/app/routers/transcription.py

@@ -0,0 +1,109 @@
+import uuid
+from typing import Optional, Union
+
+from fastapi import APIRouter, Body, Depends, Form, HTTPException, Request, UploadFile
+from pydantic import BaseModel
+from pathlib import Path
+from ..auth import apikey_auth
+from ..config import SUPPORTED_FILE_EXTENSIONS, UPLOADS_PATH
+from ..services.transcriber import MODEL_NAME
+from ..utils import cleanup_uploaded_files, download_audio_file
+
+router = APIRouter(prefix="/v1/audio", tags=["transcription"])
+
+
+class WordTiming(BaseModel):
+    word: str
+    start: float
+    end: float
+
+
+class TranscriptResult(BaseModel):
+    text: str
+    words: list[WordTiming]
+    filename: Optional[str] = None
+
+
+class TranscriptBatchResponse(BaseModel):
+    results: list[TranscriptResult]
+
+
+@router.post(
+    "/transcriptions",
+    dependencies=[Depends(apikey_auth)],
+    response_model=Union[TranscriptResult, TranscriptBatchResponse],
+)
+def transcribe(
+    request: Request,
+    file: UploadFile = None,
+    files: list[UploadFile] | None = None,
+    model: str = Form(MODEL_NAME),
+    language: str = Form("en"),
+    batch: bool = Form(False),
+):
+    service = request.app.state.whisper
+    if not file and not files:
+        raise HTTPException(
+            status_code=400, detail="Either 'file' or 'files' parameter is required"
+        )
+    if batch and not files:
+        raise HTTPException(
+            status_code=400, detail="Batch transcription requires 'files'"
+        )
+
+    upload_files = [file] if file else files
+
+    uploaded_paths: list[Path] = []
+    with cleanup_uploaded_files(uploaded_paths):
+        for upload_file in upload_files:
+            audio_suffix = upload_file.filename.split(".")[-1].lower()
+            if audio_suffix not in SUPPORTED_FILE_EXTENSIONS:
+                raise HTTPException(
+                    status_code=400,
+                    detail=(
+                        f"Unsupported audio format. Supported extensions: {', '.join(SUPPORTED_FILE_EXTENSIONS)}"
+                    ),
+                )
+            unique_filename = f"{uuid.uuid4()}.{audio_suffix}"
+            file_path = UPLOADS_PATH / unique_filename
+            with open(file_path, "wb") as f:
+                content = upload_file.file.read()
+                f.write(content)
+            uploaded_paths.append(file_path)
+
+        if batch and len(upload_files) > 1:
+            results = []
+            for path in uploaded_paths:
+                result = service.transcribe_file(str(path), language=language)
+                result["filename"] = path.name
+                results.append(result)
+            return {"results": results}
+
+        results = []
+        for path in uploaded_paths:
+            result = service.transcribe_file(str(path), language=language)
+            result["filename"] = path.name
+            results.append(result)
+
+        return {"results": results} if len(results) > 1 else results[0]
+
+
+@router.post(
+    "/transcriptions-from-url",
+    dependencies=[Depends(apikey_auth)],
+    response_model=TranscriptResult,
+)
+def transcribe_from_url(
+    request: Request,
+    audio_file_url: str = Body(..., description="URL of the audio file to transcribe"),
+    model: str = Body(MODEL_NAME),
+    language: str = Body("en"),
+    timestamp_offset: float = Body(0.0),
+):
+    service = request.app.state.whisper
+    with download_audio_file(audio_file_url) as (file_path, _ext):
+        file_path = str(file_path)
+        result = service.transcribe_vad_url_segment(
+            file_path=file_path, timestamp_offset=timestamp_offset, language=language
+        )
+        return result

gpu/self_hosted/app/routers/translation.py

@@ -0,0 +1,28 @@
+from typing import Dict
+
+from fastapi import APIRouter, Body, Depends
+from pydantic import BaseModel
+
+from ..auth import apikey_auth
+from ..services.translator import TextTranslatorService
+
+router = APIRouter(tags=["translation"])
+
+translator = TextTranslatorService()
+
+
+class TranslationResponse(BaseModel):
+    text: Dict[str, str]
+
+
+@router.post(
+    "/translate",
+    dependencies=[Depends(apikey_auth)],
+    response_model=TranslationResponse,
+)
+def translate(
+    text: str,
+    source_language: str = Body("en"),
+    target_language: str = Body("fr"),
+):
+    return translator.translate(text, source_language, target_language)

gpu/self_hosted/app/services/diarizer.py

@@ -0,0 +1,42 @@
+import os
+import threading
+
+import torch
+import torchaudio
+from pyannote.audio import Pipeline
+
+
+class PyannoteDiarizationService:
+    def __init__(self):
+        self._pipeline = None
+        self._device = "cpu"
+        self._lock = threading.Lock()
+
+    def load(self):
+        self._device = "cuda" if torch.cuda.is_available() else "cpu"
+        self._pipeline = Pipeline.from_pretrained(
+            "pyannote/speaker-diarization-3.1",
+            use_auth_token=os.environ.get("HF_TOKEN"),
+        )
+        self._pipeline.to(torch.device(self._device))
+
+    def diarize_file(self, file_path: str, timestamp: float = 0.0) -> dict:
+        if self._pipeline is None:
+            self.load()
+        waveform, sample_rate = torchaudio.load(file_path)
+        with self._lock:
+            diarization = self._pipeline(
+                {"waveform": waveform, "sample_rate": sample_rate}
+            )
+        words = []
+        for diarization_segment, _, speaker in diarization.itertracks(yield_label=True):
+            words.append(
+                {
+                    "start": round(timestamp + diarization_segment.start, 3),
+                    "end": round(timestamp + diarization_segment.end, 3),
+                    "speaker": int(speaker[-2:])
+                    if speaker and speaker[-2:].isdigit()
+                    else 0,
+                }
+            )
+        return {"diarization": words}

gpu/self_hosted/app/services/transcriber.py

@@ -0,0 +1,208 @@
+import os
+import shutil
+import subprocess
+import threading
+from typing import Generator
+
+import faster_whisper
+import librosa
+import numpy as np
+import torch
+from fastapi import HTTPException
+from silero_vad import VADIterator, load_silero_vad
+
+from ..config import SAMPLE_RATE, VAD_CONFIG
+
+# Whisper configuration (service-local defaults)
+MODEL_NAME = "large-v2"
+# None delegates compute type to runtime: float16 on CUDA, int8 on CPU
+MODEL_COMPUTE_TYPE = None
+MODEL_NUM_WORKERS = 1
+CACHE_PATH = os.path.join(os.path.expanduser("~"), ".cache", "reflector-whisper")
+from ..utils import NoStdStreams
+
+
+class WhisperService:
+    def __init__(self):
+        self.model = None
+        self.device = "cpu"
+        self.lock = threading.Lock()
+
+    def load(self):
+        self.device = "cuda" if torch.cuda.is_available() else "cpu"
+        compute_type = MODEL_COMPUTE_TYPE or (
+            "float16" if self.device == "cuda" else "int8"
+        )
+        self.model = faster_whisper.WhisperModel(
+            MODEL_NAME,
+            device=self.device,
+            compute_type=compute_type,
+            num_workers=MODEL_NUM_WORKERS,
+            download_root=CACHE_PATH,
+        )
+
+    def pad_audio(self, audio_array, sample_rate: int = SAMPLE_RATE):
+        audio_duration = len(audio_array) / sample_rate
+        if audio_duration < VAD_CONFIG["silence_padding"]:
+            silence_samples = int(sample_rate * VAD_CONFIG["silence_padding"])
+            silence = np.zeros(silence_samples, dtype=np.float32)
+            return np.concatenate([audio_array, silence])
+        return audio_array
+
+    def enforce_word_timing_constraints(self, words: list[dict]) -> list[dict]:
+        if len(words) <= 1:
+            return words
+        enforced: list[dict] = []
+        for i, word in enumerate(words):
+            current = dict(word)
+            if i < len(words) - 1:
+                next_start = words[i + 1]["start"]
+                if current["end"] > next_start:
+                    current["end"] = next_start
+            enforced.append(current)
+        return enforced
+
+    def transcribe_file(self, file_path: str, language: str = "en") -> dict:
+        input_for_model: str | "object" = file_path
+        try:
+            audio_array, _sample_rate = librosa.load(
+                file_path, sr=SAMPLE_RATE, mono=True
+            )
+            if len(audio_array) / float(SAMPLE_RATE) < VAD_CONFIG["silence_padding"]:
+                input_for_model = self.pad_audio(audio_array, SAMPLE_RATE)
+        except Exception:
+            pass
+
+        with self.lock:
+            with NoStdStreams():
+                segments, _ = self.model.transcribe(
+                    input_for_model,
+                    language=language,
+                    beam_size=5,
+                    word_timestamps=True,
+                    vad_filter=True,
+                    vad_parameters={"min_silence_duration_ms": 500},
+                )
+
+        segments = list(segments)
+        text = "".join(segment.text for segment in segments).strip()
+        words = [
+            {
+                "word": word.word,
+                "start": round(float(word.start), 2),
+                "end": round(float(word.end), 2),
+            }
+            for segment in segments
+            for word in segment.words
+        ]
+        words = self.enforce_word_timing_constraints(words)
+        return {"text": text, "words": words}
+
+    def transcribe_vad_url_segment(
+        self, file_path: str, timestamp_offset: float = 0.0, language: str = "en"
+    ) -> dict:
+        def load_audio_via_ffmpeg(input_path: str, sample_rate: int) -> np.ndarray:
+            ffmpeg_bin = shutil.which("ffmpeg") or "ffmpeg"
+            cmd = [
+                ffmpeg_bin,
+                "-nostdin",
+                "-threads",
+                "1",
+                "-i",
+                input_path,
+                "-f",
+                "f32le",
+                "-acodec",
+                "pcm_f32le",
+                "-ac",
+                "1",
+                "-ar",
+                str(sample_rate),
+                "pipe:1",
+            ]
+            try:
+                proc = subprocess.run(
+                    cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE, check=True
+                )
+            except Exception as e:
+                raise HTTPException(status_code=400, detail=f"ffmpeg failed: {e}")
+            audio = np.frombuffer(proc.stdout, dtype=np.float32)
+            return audio
+
+        def vad_segments(
+            audio_array,
+            sample_rate: int = SAMPLE_RATE,
+            window_size: int = VAD_CONFIG["window_size"],
+        ) -> Generator[tuple[float, float], None, None]:
+            vad_model = load_silero_vad(onnx=False)
+            iterator = VADIterator(vad_model, sampling_rate=sample_rate)
+            start = None
+            for i in range(0, len(audio_array), window_size):
+                chunk = audio_array[i : i + window_size]
+                if len(chunk) < window_size:
+                    chunk = np.pad(
+                        chunk, (0, window_size - len(chunk)), mode="constant"
+                    )
+                speech = iterator(chunk)
+                if not speech:
+                    continue
+                if "start" in speech:
+                    start = speech["start"]
+                    continue
+                if "end" in speech and start is not None:
+                    end = speech["end"]
+                    yield (start / float(SAMPLE_RATE), end / float(SAMPLE_RATE))
+                    start = None
+            iterator.reset_states()
+
+        audio_array = load_audio_via_ffmpeg(file_path, SAMPLE_RATE)
+
+        merged_batches: list[tuple[float, float]] = []
+        batch_start = None
+        batch_end = None
+        max_duration = VAD_CONFIG["batch_max_duration"]
+        for seg_start, seg_end in vad_segments(audio_array):
+            if batch_start is None:
+                batch_start, batch_end = seg_start, seg_end
+                continue
+            if seg_end - batch_start <= max_duration:
+                batch_end = seg_end
+            else:
+                merged_batches.append((batch_start, batch_end))
+                batch_start, batch_end = seg_start, seg_end
+        if batch_start is not None and batch_end is not None:
+            merged_batches.append((batch_start, batch_end))
+
+        all_text = []
+        all_words = []
+        for start_time, end_time in merged_batches:
+            s_idx = int(start_time * SAMPLE_RATE)
+            e_idx = int(end_time * SAMPLE_RATE)
+            segment = audio_array[s_idx:e_idx]
+            segment = self.pad_audio(segment, SAMPLE_RATE)
+            with self.lock:
+                segments, _ = self.model.transcribe(
+                    segment,
+                    language=language,
+                    beam_size=5,
+                    word_timestamps=True,
+                    vad_filter=True,
+                    vad_parameters={"min_silence_duration_ms": 500},
+                )
+            segments = list(segments)
+            text = "".join(seg.text for seg in segments).strip()
+            words = [
+                {
+                    "word": w.word,
+                    "start": round(float(w.start) + start_time + timestamp_offset, 2),
+                    "end": round(float(w.end) + start_time + timestamp_offset, 2),
+                }
+                for seg in segments
+                for w in seg.words
+            ]
+            if text:
+                all_text.append(text)
+            all_words.extend(words)
+
+        all_words = self.enforce_word_timing_constraints(all_words)
+        return {"text": " ".join(all_text), "words": all_words}

gpu/self_hosted/app/services/translator.py

@@ -0,0 +1,44 @@
+import threading
+
+from transformers import MarianMTModel, MarianTokenizer, pipeline
+
+
+class TextTranslatorService:
+    """Simple text-to-text translator using HuggingFace MarianMT models.
+
+    This mirrors the modal translator API shape but uses text translation only.
+    """
+
+    def __init__(self):
+        self._pipeline = None
+        self._lock = threading.Lock()
+
+    def load(self, source_language: str = "en", target_language: str = "fr"):
+        # Pick a default MarianMT model pair if available; fall back to Helsinki-NLP en->fr
+        model_name = self._resolve_model_name(source_language, target_language)
+        tokenizer = MarianTokenizer.from_pretrained(model_name)
+        model = MarianMTModel.from_pretrained(model_name)
+        self._pipeline = pipeline("translation", model=model, tokenizer=tokenizer)
+
+    def _resolve_model_name(self, src: str, tgt: str) -> str:
+        # Minimal mapping; extend as needed
+        pair = (src.lower(), tgt.lower())
+        mapping = {
+            ("en", "fr"): "Helsinki-NLP/opus-mt-en-fr",
+            ("fr", "en"): "Helsinki-NLP/opus-mt-fr-en",
+            ("en", "es"): "Helsinki-NLP/opus-mt-en-es",
+            ("es", "en"): "Helsinki-NLP/opus-mt-es-en",
+            ("en", "de"): "Helsinki-NLP/opus-mt-en-de",
+            ("de", "en"): "Helsinki-NLP/opus-mt-de-en",
+        }
+        return mapping.get(pair, "Helsinki-NLP/opus-mt-en-fr")
+
+    def translate(self, text: str, source_language: str, target_language: str) -> dict:
+        if self._pipeline is None:
+            self.load(source_language, target_language)
+        with self._lock:
+            results = self._pipeline(
+                text, src_lang=source_language, tgt_lang=target_language
+            )
+        translated = results[0]["translation_text"] if results else ""
+        return {"text": {source_language: text, target_language: translated}}

gpu/self_hosted/app/utils.py

@@ -0,0 +1,107 @@
+import logging
+import os
+import sys
+import uuid
+from contextlib import contextmanager
+from typing import Mapping
+from urllib.parse import urlparse
+from pathlib import Path
+
+import requests
+from fastapi import HTTPException
+
+from .config import SUPPORTED_FILE_EXTENSIONS, UPLOADS_PATH
+
+logger = logging.getLogger(__name__)
+
+
+class NoStdStreams:
+    def __init__(self):
+        self.devnull = open(os.devnull, "w")
+
+    def __enter__(self):
+        self._stdout, self._stderr = sys.stdout, sys.stderr
+        self._stdout.flush()
+        self._stderr.flush()
+        sys.stdout, sys.stderr = self.devnull, self.devnull
+
+    def __exit__(self, exc_type, exc_value, traceback):
+        sys.stdout, sys.stderr = self._stdout, self._stderr
+        self.devnull.close()
+
+
+def ensure_dirs():
+    UPLOADS_PATH.mkdir(parents=True, exist_ok=True)
+
+
+def detect_audio_format(url: str, headers: Mapping[str, str]) -> str:
+    url_path = urlparse(url).path
+    for ext in SUPPORTED_FILE_EXTENSIONS:
+        if url_path.lower().endswith(f".{ext}"):
+            return ext
+
+    content_type = headers.get("content-type", "").lower()
+    if "audio/mpeg" in content_type or "audio/mp3" in content_type:
+        return "mp3"
+    if "audio/wav" in content_type:
+        return "wav"
+    if "audio/mp4" in content_type:
+        return "mp4"
+
+    raise HTTPException(
+        status_code=400,
+        detail=(
+            f"Unsupported audio format for URL. Supported extensions: {', '.join(SUPPORTED_FILE_EXTENSIONS)}"
+        ),
+    )
+
+
+def download_audio_to_uploads(audio_file_url: str) -> tuple[Path, str]:
+    response = requests.head(audio_file_url, allow_redirects=True)
+    if response.status_code == 404:
+        raise HTTPException(status_code=404, detail="Audio file not found")
+
+    response = requests.get(audio_file_url, allow_redirects=True)
+    response.raise_for_status()
+
+    audio_suffix = detect_audio_format(audio_file_url, response.headers)
+    unique_filename = f"{uuid.uuid4()}.{audio_suffix}"
+    file_path: Path = UPLOADS_PATH / unique_filename
+
+    with open(file_path, "wb") as f:
+        f.write(response.content)
+
+    return file_path, audio_suffix
+
+
+@contextmanager
+def download_audio_file(audio_file_url: str):
+    """Download an audio file to UPLOADS_PATH and remove it after use.
+
+    Yields (file_path: Path, audio_suffix: str).
+    """
+    file_path, audio_suffix = download_audio_to_uploads(audio_file_url)
+    try:
+        yield file_path, audio_suffix
+    finally:
+        try:
+            file_path.unlink(missing_ok=True)
+        except Exception as e:
+            logger.error("Error deleting temporary file %s: %s", file_path, e)
+
+
+@contextmanager
+def cleanup_uploaded_files(file_paths: list[Path]):
+    """Ensure provided file paths are removed after use.
+
+    The provided list can be populated inside the context; all present entries
+    at exit will be deleted.
+    """
+    try:
+        yield file_paths
+    finally:
+        for path in list(file_paths):
+            try:
+                path.unlink(missing_ok=True)
+            except Exception as e:
+                logger.error("Error deleting temporary file %s: %s", path, e)

gpu/self_hosted/compose.yml

@@ -0,0 +1,10 @@
+services:
+  reflector_gpu:
+    build:
+      context: .
+    ports:
+      - "8000:8000"
+    env_file:
+      - .env
+    volumes:
+      - ./cache:/root/.cache

gpu/self_hosted/main.py

@@ -0,0 +1,3 @@
+from app.factory import create_app
+
+app = create_app()

gpu/self_hosted/pyproject.toml

@@ -0,0 +1,19 @@
+[project]
+name = "reflector-gpu"
+version = "0.1.0"
+description = "Self-hosted GPU service for speech transcription, diarization, and translation via FastAPI."
+readme = "README.md"
+requires-python = ">=3.12"
+dependencies = [
+    "fastapi[standard]>=0.116.1",
+    "uvicorn[standard]>=0.30.0",
+    "torch>=2.3.0",
+    "faster-whisper>=1.1.0",
+    "librosa==0.10.1",
+    "numpy<2",
+    "silero-vad==5.1.0",
+    "transformers>=4.35.0",
+    "sentencepiece",
+    "pyannote.audio==3.1.0",
+    "torchaudio>=2.3.0",
+]

gpu/self_hosted/runserver.sh

@@ -0,0 +1,17 @@
+#!/bin/sh
+set -e
+
+export PATH="/root/.local/bin:$PATH"
+cd /app
+
+# Install Python dependencies at runtime (first run or when FORCE_SYNC=1)
+if [ ! -d "/app/.venv" ] || [ "$FORCE_SYNC" = "1" ]; then
+  echo "[startup] Installing Python dependencies with uv..."
+  uv sync --compile-bytecode --locked
+else
+  echo "[startup] Using existing virtual environment at /app/.venv"
+fi
+
+exec uv run uvicorn main:app --host 0.0.0.0 --port 8000
+
+

server/DAILYCO_TEST.md

@@ -0,0 +1,613 @@
+# Daily.co Integration Test Plan
+
+## ✅ IMPLEMENTATION STATUS: Real Transcription Active
+
+**This test validates Daily.co multitrack recording integration with REAL transcription/diarization.**
+
+The implementation includes complete audio processing pipeline:
+- **Multitrack recordings** from Daily.co S3 (separate audio stream per participant)
+- **PyAV-based audio mixdown** with PTS-based track alignment
+- **Real transcription** via Modal GPU backend (Whisper)
+- **Real diarization** via Modal GPU backend (speaker identification)
+- **Per-track transcription** with timestamp synchronization
+- **Complete database entities** (recording, transcript, topics, participants, words)
+
+**Processing pipeline** (`PipelineMainMultitrack`):
+1. Download all audio tracks from Daily.co S3
+2. Align tracks by PTS (presentation timestamp) to handle late joiners
+3. Mix tracks into single audio file for unified playback
+4. Transcribe each track individually with proper offset handling
+5. Perform diarization on mixed audio
+6. Generate topics, summaries, and word-level timestamps
+7. Convert audio to MP3 and generate waveform visualization
+
+**Note:** A stub processor (`process_daily_recording`) exists for testing webhook flow without GPU costs, but the production code path uses `process_multitrack_recording` with full ML pipeline.
+
+---
+
+## Prerequisites
+
+**1. Environment Variables** (check in `.env.development.local`):
+```bash
+# Daily.co API Configuration
+DAILY_API_KEY=<key>
+DAILY_SUBDOMAIN=monadical
+DAILY_WEBHOOK_SECRET=<base64-encoded-secret>
+AWS_DAILY_S3_BUCKET=reflector-dailyco-local
+AWS_DAILY_S3_REGION=us-east-1
+AWS_DAILY_ROLE_ARN=arn:aws:iam::950402358378:role/DailyCo
+DAILY_MIGRATION_ENABLED=true
+DAILY_MIGRATION_ROOM_IDS=["552640fd-16f2-4162-9526-8cf40cd2357e"]
+
+# Transcription/Diarization Backend (Required for real processing)
+DIARIZATION_BACKEND=modal
+DIARIZATION_MODAL_API_KEY=<modal-api-key>
+# TRANSCRIPTION_BACKEND is not explicitly set (uses default/modal)
+```
+
+**2. Services Running:**
+```bash
+docker compose ps  # server, postgres, redis, worker, beat should be UP
+```
+
+**IMPORTANT:** Worker and beat services MUST be running for transcription processing:
+```bash
+docker compose up -d worker beat
+```
+
+**3. ngrok Tunnel for Webhooks:**
+```bash
+# Start ngrok (if not already running)
+ngrok http 1250 --log=stdout > /tmp/ngrok.log 2>&1 &
+
+# Get public URL
+curl -s http://localhost:4040/api/tunnels | python3 -c "import sys, json; data=json.load(sys.stdin); print(data['tunnels'][0]['public_url'])"
+```
+
+**Current ngrok URL:** `https://0503947384a3.ngrok-free.app` (as of last registration)
+
+**4. Webhook Created:**
+```bash
+cd server
+uv run python scripts/recreate_daily_webhook.py https://0503947384a3.ngrok-free.app/v1/daily/webhook
+# Verify: "Created webhook <uuid> (state: ACTIVE)"
+```
+
+**Current webhook status:** ✅ ACTIVE (webhook ID: dad5ad16-ceca-488e-8fc5-dae8650b51d0)
+
+---
+
+## Test 1: Database Configuration
+
+**Check room platform:**
+```bash
+docker-compose exec -T postgres psql -U reflector -d reflector -c \
+  "SELECT id, name, platform, recording_type FROM room WHERE name = 'test2';"
+```
+
+**Expected:**
+```
+id: 552640fd-16f2-4162-9526-8cf40cd2357e
+name: test2
+platform: whereby  # DB value (overridden by env var DAILY_MIGRATION_ROOM_IDS)
+recording_type: cloud
+```
+
+**Clear old meetings:**
+```bash
+docker-compose exec -T postgres psql -U reflector -d reflector -c \
+  "UPDATE meeting SET is_active = false WHERE room_id = '552640fd-16f2-4162-9526-8cf40cd2357e';"
+```
+
+---
+
+## Test 2: Meeting Creation with Auto-Recording
+
+**Create meeting:**
+```bash
+curl -s -X POST http://localhost:1250/v1/rooms/test2/meeting \
+  -H "Content-Type: application/json" \
+  -d '{"allow_duplicated":false}' | python3 -m json.tool
+```
+
+**Expected Response:**
+```json
+{
+  "room_name": "test2-YYYYMMDDHHMMSS",  // Includes "test2" prefix!
+  "room_url": "https://monadical.daily.co/test2-...?t=<JWT_TOKEN>",  // Has token!
+  "platform": "daily",
+  "recording_type": "cloud"  // DB value (Whereby-specific)
+}
+```
+
+**Decode token to verify auto-recording:**
+```bash
+# Extract token from room_url, decode JWT payload
+echo "<token>" | python3 -c "
+import sys, json, base64
+token = sys.stdin.read().strip()
+payload = token.split('.')[1] + '=' * (4 - len(token.split('.')[1]) % 4)
+print(json.dumps(json.loads(base64.b64decode(payload)), indent=2))
+"
+```
+
+**Expected token payload:**
+```json
+{
+  "r": "test2-YYYYMMDDHHMMSS",  // Room name
+  "sr": true,  // start_recording: true ✅
+  "d": "...",  // Domain ID
+  "iat": 1234567890
+}
+```
+
+---
+
+## Test 3: Daily.co API Verification
+
+**Check room configuration:**
+```bash
+ROOM_NAME="<from previous step>"
+curl -s -X GET "https://api.daily.co/v1/rooms/$ROOM_NAME" \
+  -H "Authorization: Bearer $DAILY_API_KEY" | python3 -m json.tool
+```
+
+**Expected config:**
+```json
+{
+  "config": {
+    "enable_recording": "raw-tracks",  // ✅
+    "recordings_bucket": {
+      "bucket_name": "reflector-dailyco-local",
+      "bucket_region": "us-east-1",
+      "assume_role_arn": "arn:aws:iam::950402358378:role/DailyCo"
+    }
+  }
+}
+```
+
+---
+
+## Test 4: Browser UI Test (Playwright MCP)
+
+**Using Claude Code MCP tools:**
+
+**Load room:**
+```
+Use: mcp__playwright__browser_navigate
+Input: {"url": "http://localhost:3000/test2"}
+
+Then wait 12 seconds for iframe to load
+```
+
+**Verify Daily.co iframe loaded:**
+```
+Use: mcp__playwright__browser_snapshot
+
+Expected in snapshot:
+- iframe element with src containing "monadical.daily.co"
+- Daily.co pre-call UI visible
+```
+
+**Take screenshot:**
+```
+Use: mcp__playwright__browser_take_screenshot
+Input: {"filename": "test2-before-join.png"}
+
+Expected: Daily.co pre-call UI with "Join" button visible
+```
+
+**Join meeting:**
+```
+Note: Daily.co iframe interaction requires clicking inside iframe.
+Use: mcp__playwright__browser_click
+Input: {"element": "Join button in Daily.co iframe", "ref": "<ref-from-snapshot>"}
+
+Then wait 5 seconds for call to connect
+```
+
+**Verify in-call:**
+```
+Use: mcp__playwright__browser_take_screenshot
+Input: {"filename": "test2-in-call.png"}
+
+Expected: "Waiting for others to join" or participant video visible
+```
+
+**Leave meeting:**
+```
+Use: mcp__playwright__browser_click
+Input: {"element": "Leave button in Daily.co iframe", "ref": "<ref-from-snapshot>"}
+```
+
+---
+
+**Alternative: JavaScript snippets (for manual testing):**
+
+```javascript
+await page.goto('http://localhost:3000/test2');
+await new Promise(f => setTimeout(f, 12000));  // Wait for load
+
+// Verify iframe
+const iframes = document.querySelectorAll('iframe');
+// Expected: 1 iframe with src containing "monadical.daily.co"
+
+// Screenshot
+await page.screenshot({ path: 'test2-before-join.png' });
+
+// Join
+await page.locator('iframe').contentFrame().getByRole('button', { name: 'Join' }).click();
+await new Promise(f => setTimeout(f, 5000));
+
+// In-call screenshot
+await page.screenshot({ path: 'test2-in-call.png' });
+
+// Leave
+await page.locator('iframe').contentFrame().getByRole('button', { name: 'Leave' }).click();
+```
+
+---
+
+## Test 5: Webhook Verification
+
+**Check server logs for webhooks:**
+```bash
+docker-compose logs --since 15m server 2>&1 | grep -i "participant joined\|recording started"
+```
+
+**Expected logs:**
+```
+[info] Participant joined | meeting_id=... | num_clients=1 | recording_type=cloud | recording_trigger=automatic-2nd-participant
+[info] Recording started | meeting_id=... | recording_id=... | platform=daily
+```
+
+**Check Daily.co webhook delivery logs:**
+```bash
+curl -s -X GET "https://api.daily.co/v1/logs/webhooks?limit=20" \
+  -H "Authorization: Bearer $DAILY_API_KEY" | python3 -c "
+import sys, json
+logs = json.load(sys.stdin)
+for log in logs[:10]:
+    req = json.loads(log['request'])
+    room = req.get('payload', {}).get('room') or req.get('payload', {}).get('room_name', 'N/A')
+    print(f\"{req['type']:30s} | room: {room:30s} | status: {log['status']}\")
+"
+```
+
+**Expected output:**
+```
+participant.joined             | room: test2-YYYYMMDDHHMMSS       | status: 200
+recording.started              | room: test2-YYYYMMDDHHMMSS       | status: 200
+participant.left               | room: test2-YYYYMMDDHHMMSS       | status: 200
+recording.ready-to-download    | room: test2-YYYYMMDDHHMMSS       | status: 200
+```
+
+**Check database updated:**
+```bash
+docker-compose exec -T postgres psql -U reflector -d reflector -c \
+  "SELECT room_name, num_clients FROM meeting WHERE room_name LIKE 'test2-%' ORDER BY end_date DESC LIMIT 1;"
+```
+
+**Expected:**
+```
+room_name: test2-YYYYMMDDHHMMSS
+num_clients: 0  // After participant left
+```
+
+---
+
+## Test 6: Recording in S3
+
+**List recent recordings:**
+```bash
+curl -s -X GET "https://api.daily.co/v1/recordings" \
+  -H "Authorization: Bearer $DAILY_API_KEY" | python3 -c "
+import sys, json
+data = json.load(sys.stdin)
+for rec in data.get('data', [])[:5]:
+    if 'test2-' in rec.get('room_name', ''):
+        print(f\"Room: {rec['room_name']}\")
+        print(f\"Status: {rec['status']}\")
+        print(f\"Duration: {rec.get('duration', 0)}s\")
+        print(f\"S3 key: {rec.get('s3key', 'N/A')}\")
+        print(f\"Tracks: {len(rec.get('tracks', []))} files\")
+        for track in rec.get('tracks', []):
+            print(f\"  - {track['type']}: {track['s3Key'].split('/')[-1]} ({track['size']} bytes)\")
+        print()
+"
+```
+
+**Expected output:**
+```
+Room: test2-20251009192341
+Status: finished
+Duration: ~30-120s
+S3 key: monadical/test2-20251009192341/1760037914930
+Tracks: 2 files
+  - audio: 1760037914930-<uuid>-cam-audio-1760037915265 (~400 KB)
+  - video: 1760037914930-<uuid>-cam-video-1760037915269 (~10-30 MB)
+```
+
+**Verify S3 path structure:**
+- `monadical/` - Daily.co subdomain
+- `test2-20251009192341/` - Reflector room name + timestamp
+- `<timestamp>-<participant-uuid>-<media-type>-<track-start>.webm` - Individual track files
+
+---
+
+## Test 7: Database Check - Recording and Transcript
+
+**Check recording created:**
+```bash
+docker-compose exec -T postgres psql -U reflector -d reflector -c \
+  "SELECT id, bucket_name, object_key, status, meeting_id, recorded_at
+   FROM recording
+   ORDER BY recorded_at DESC LIMIT 1;"
+```
+
+**Expected:**
+```
+id: <recording-id-from-webhook>
+bucket_name: reflector-dailyco-local
+object_key: monadical/test2-<timestamp>/<recording-timestamp>-<uuid>-cam-audio-<track-start>.webm
+status: completed
+meeting_id: <meeting-id>
+recorded_at: <recent-timestamp>
+```
+
+**Check transcript created:**
+```bash
+docker compose exec -T postgres psql -U reflector -d reflector -c \
+  "SELECT id, title, status, duration, recording_id, meeting_id, room_id
+   FROM transcript
+   ORDER BY created_at DESC LIMIT 1;"
+```
+
+**Expected (REAL transcription):**
+```
+id: <transcript-id>
+title: <AI-generated title based on actual conversation content>
+status: uploaded  (audio file processed and available)
+duration: <actual meeting duration in seconds>
+recording_id: <same-as-recording-id-above>
+meeting_id: <meeting-id>
+room_id: 552640fd-16f2-4162-9526-8cf40cd2357e
+```
+
+**Note:** Title and content will reflect the ACTUAL conversation, not mock data. Processing time depends on recording length and GPU backend availability (Modal).
+
+**Verify audio file exists:**
+```bash
+ls -lh data/<transcript-id>/upload.webm
+```
+
+**Expected:**
+```
+-rw-r--r-- 1 user staff ~100-200K Oct 10 18:48 upload.webm
+```
+
+**Check transcript topics (REAL transcription):**
+```bash
+TRANSCRIPT_ID=$(docker compose exec -T postgres psql -U reflector -d reflector -t -c \
+  "SELECT id FROM transcript ORDER BY created_at DESC LIMIT 1;")
+
+docker compose exec -T postgres psql -U reflector -d reflector -c \
+  "SELECT
+     jsonb_array_length(topics) as num_topics,
+     jsonb_array_length(participants) as num_participants,
+     short_summary,
+     title
+   FROM transcript
+   WHERE id = '$TRANSCRIPT_ID';"
+```
+
+**Expected (REAL data):**
+```
+num_topics: <varies based on conversation>
+num_participants: <actual number of participants who spoke>
+short_summary: <AI-generated summary of actual conversation>
+title: <AI-generated title based on content>
+```
+
+**Check topics contain actual transcription:**
+```bash
+docker compose exec -T postgres psql -U reflector -d reflector -c \
+  "SELECT topics->0->'title', topics->0->'summary', topics->0->'transcript'
+   FROM transcript
+   ORDER BY created_at DESC LIMIT 1;" | head -20
+```
+
+**Expected output:** Will contain the ACTUAL transcribed conversation from the Daily.co meeting, not mock data.
+
+**Check participants:**
+```bash
+docker compose exec -T postgres psql -U reflector -d reflector -c \
+  "SELECT participants FROM transcript ORDER BY created_at DESC LIMIT 1;" \
+  | python3 -c "import sys, json; data=json.loads(sys.stdin.read()); print(json.dumps(data, indent=2))"
+```
+
+**Expected (REAL diarization):**
+```json
+[
+  {
+    "id": "<uuid>",
+    "speaker": 0,
+    "name": "Speaker 1"
+  },
+  {
+    "id": "<uuid>",
+    "speaker": 1,
+    "name": "Speaker 2"
+  }
+]
+```
+
+**Note:** Speaker names will be generic ("Speaker 1", "Speaker 2", etc.) as determined by the diarization backend. Number of participants depends on how many actually spoke during the meeting.
+
+**Check word-level data:**
+```bash
+docker compose exec -T postgres psql -U reflector -d reflector -c \
+  "SELECT jsonb_array_length(topics->0->'words') as num_words_first_topic
+   FROM transcript
+   ORDER BY created_at DESC LIMIT 1;"
+```
+
+**Expected:**
+```
+num_words_first_topic: <varies based on actual conversation length and topic chunking>
+```
+
+**Verify speaker diarization in words:**
+```bash
+docker compose exec -T postgres psql -U reflector -d reflector -c \
+  "SELECT
+     topics->0->'words'->0->>'text' as first_word,
+     topics->0->'words'->0->>'speaker' as speaker,
+     topics->0->'words'->0->>'start' as start_time,
+     topics->0->'words'->0->>'end' as end_time
+   FROM transcript
+   ORDER BY created_at DESC LIMIT 1;"
+```
+
+**Expected (REAL transcription):**
+```
+first_word: <actual first word from transcription>
+speaker: 0, 1, 2, ... (actual speaker ID from diarization)
+start_time: <actual timestamp in seconds>
+end_time: <actual end timestamp>
+```
+
+**Note:** All timestamps and speaker IDs are from real transcription/diarization, synchronized across tracks.
+
+---
+
+## Test 8: Recording Type Verification
+
+**Check what Daily.co received:**
+```bash
+curl -s -X GET "https://api.daily.co/v1/rooms/test2-<timestamp>" \
+  -H "Authorization: Bearer $DAILY_API_KEY" | python3 -m json.tool | grep "enable_recording"
+```
+
+**Expected:**
+```json
+"enable_recording": "raw-tracks"
+```
+
+**NOT:** `"enable_recording": "cloud"` (that would be wrong - we want raw tracks)
+
+---
+
+## Troubleshooting
+
+### Issue: No webhooks received
+
+**Check webhook state:**
+```bash
+curl -s -X GET "https://api.daily.co/v1/webhooks" \
+  -H "Authorization: Bearer $DAILY_API_KEY" | python3 -m json.tool
+```
+
+**If state is FAILED:**
+```bash
+cd server
+uv run python scripts/recreate_daily_webhook.py https://<ngrok-url>/v1/daily/webhook
+```
+
+### Issue: Webhooks return 422
+
+**Check server logs:**
+```bash
+docker-compose logs --tail=50 server | grep "Failed to parse webhook event"
+```
+
+**Common cause:** Event structure mismatch. Daily.co events use:
+```json
+{
+  "version": "1.0.0",
+  "type": "participant.joined",
+  "payload": {...},  // NOT "data"
+  "event_ts": 123.456  // NOT "ts"
+}
+```
+
+### Issue: Recording not starting
+
+1. **Check token has `sr: true`:**
+   - Decode JWT token from room_url query param
+   - Should contain `"sr": true`
+
+2. **Check Daily.co room config:**
+   - `enable_recording` must be set (not false)
+   - For raw-tracks: must be exactly `"raw-tracks"`
+
+3. **Check participant actually joined:**
+   - Logs should show "Participant joined"
+   - Must click "Join" button, not just pre-call screen
+
+### Issue: Recording in S3 but wrong format
+
+**Daily.co recording types:**
+- `"cloud"` → Single MP4 file (`download_link` in webhook)
+- `"raw-tracks"` → Multiple WebM files (`tracks` array in webhook)
+- `"raw-tracks-audio-only"` → Only audio WebM files
+
+**Current implementation:** Always uses `"raw-tracks"` (better for transcription)
+
+---
+
+## Quick Validation Commands
+
+**One-liner to verify everything:**
+```bash
+# 1. Check room exists
+docker-compose exec -T postgres psql -U reflector -d reflector -c \
+  "SELECT name, platform FROM room WHERE name = 'test2';" && \
+
+# 2. Create meeting
+MEETING=$(curl -s -X POST http://localhost:1250/v1/rooms/test2/meeting \
+  -H "Content-Type: application/json" -d '{"allow_duplicated":false}') && \
+echo "$MEETING" | python3 -c "import sys,json; m=json.load(sys.stdin); print(f'Room: {m[\"room_name\"]}\nURL: {m[\"room_url\"][:80]}...')" && \
+
+# 3. Check Daily.co config
+ROOM_NAME=$(echo "$MEETING" | python3 -c "import sys,json; print(json.load(sys.stdin)['room_name'])") && \
+curl -s -X GET "https://api.daily.co/v1/rooms/$ROOM_NAME" \
+  -H "Authorization: Bearer $DAILY_API_KEY" | python3 -c "import sys,json; print(f'Recording: {json.load(sys.stdin)[\"config\"][\"enable_recording\"]}')"
+```
+
+**Expected output:**
+```
+name: test2, platform: whereby
+Room: test2-20251009192341
+URL: https://monadical.daily.co/test2-20251009192341?t=eyJhbGc...
+Recording: raw-tracks
+```
+
+---
+
+## Success Criteria Checklist
+
+- [x] Room name includes Reflector room prefix (`test2-...`)
+- [x] Meeting URL contains JWT token (`?t=...`)
+- [x] Token has `sr: true` (auto-recording enabled)
+- [x] Daily.co room config: `enable_recording: "raw-tracks"`
+- [x] Browser loads Daily.co interface (not Whereby)
+- [x] Recording auto-starts when participant joins
+- [x] Webhooks received: participant.joined, recording.started, participant.left, recording.ready-to-download
+- [x] Recording status: `finished`
+- [x] S3 contains 2 files: audio (.webm) and video (.webm)
+- [x] S3 path: `monadical/test2-{timestamp}/{recording-start-ts}-{participant-uuid}-cam-{audio|video}-{track-start-ts}`
+- [x] Database `num_clients` increments/decrements correctly
+- [x] **Database recording entry created** with correct S3 path and status `completed`
+- [ ] **Database transcript entry created** with status `uploaded`
+- [ ] **Audio file downloaded** to `data/{transcript_id}/upload.webm`
+- [ ] **Transcript has REAL data**: AI-generated title based on conversation
+- [ ] **Transcript has topics** generated from actual content
+- [ ] **Transcript has participants** with proper speaker diarization
+- [ ] **Topics contain word-level data** with accurate timestamps and speaker IDs
+- [ ] **Total duration** matches actual meeting length
+- [ ] **MP3 and waveform files generated** by file processing pipeline
+- [ ] **Frontend transcript page loads** without "Failed to load audio" error
+- [ ] **Audio player functional** with working playback and waveform visualization
+- [ ] **Multitrack processing completed** without errors in worker logs
+- [ ] **Modal GPU backends accessible** (transcription and diarization)

server/Dockerfile

@@ -6,7 +6,7 @@ ENV PYTHONUNBUFFERED=1 \
 
 # builder install base dependencies
 WORKDIR /tmp
-RUN apt-get update && apt-get install -y curl && apt-get clean
+RUN apt-get update && apt-get install -y curl ffmpeg && apt-get clean
 ADD https://astral.sh/uv/install.sh /uv-installer.sh
 RUN sh /uv-installer.sh && rm /uv-installer.sh
 ENV PATH="/root/.local/bin/:$PATH"

server/docs/data_retention.md

@@ -0,0 +1,95 @@
+# Data Retention and Cleanup
+
+## Overview
+
+For public instances of Reflector, a data retention policy is automatically enforced to delete anonymous user data after a configurable period (default: 7 days). This ensures compliance with privacy expectations and prevents unbounded storage growth.
+
+## Configuration
+
+### Environment Variables
+
+- `PUBLIC_MODE` (bool): Must be set to `true` to enable automatic cleanup
+- `PUBLIC_DATA_RETENTION_DAYS` (int): Number of days to retain anonymous data (default: 7)
+
+### What Gets Deleted
+
+When data reaches the retention period, the following items are automatically removed:
+
+1. **Transcripts** from anonymous users (where `user_id` is NULL):
+   - Database records
+   - Local files (audio.wav, audio.mp3, audio.json waveform)
+   - Storage files (cloud storage if configured)
+
+## Automatic Cleanup
+
+### Celery Beat Schedule
+
+When `PUBLIC_MODE=true`, a Celery beat task runs daily at 3 AM to clean up old data:
+
+```python
+# Automatically scheduled when PUBLIC_MODE=true
+"cleanup_old_public_data": {
+    "task": "reflector.worker.cleanup.cleanup_old_public_data",
+    "schedule": crontab(hour=3, minute=0),  # Daily at 3 AM
+}
+```
+
+### Running the Worker
+
+Ensure both Celery worker and beat scheduler are running:
+
+```bash
+# Start Celery worker
+uv run celery -A reflector.worker.app worker --loglevel=info
+
+# Start Celery beat scheduler (in another terminal)
+uv run celery -A reflector.worker.app beat
+```
+
+## Manual Cleanup
+
+For testing or manual intervention, use the cleanup tool:
+
+```bash
+# Delete data older than 7 days (default)
+uv run python -m reflector.tools.cleanup_old_data
+
+# Delete data older than 30 days
+uv run python -m reflector.tools.cleanup_old_data --days 30
+```
+
+Note: The manual tool uses the same implementation as the Celery worker task to ensure consistency.
+
+## Important Notes
+
+1. **User Data Deletion**: Only anonymous data (where `user_id` is NULL) is deleted. Authenticated user data is preserved.
+
+2. **Storage Cleanup**: The system properly cleans up both local files and cloud storage when configured.
+
+3. **Error Handling**: If individual deletions fail, the cleanup continues and logs errors. Failed deletions are reported in the task output.
+
+4. **Public Instance Only**: The automatic cleanup task only runs when `PUBLIC_MODE=true` to prevent accidental data loss in private deployments.
+
+## Testing
+
+Run the cleanup tests:
+
+```bash
+uv run pytest tests/test_cleanup.py -v
+```
+
+## Monitoring
+
+Check Celery logs for cleanup task execution:
+
+```bash
+# Look for cleanup task logs
+grep "cleanup_old_public_data" celery.log
+grep "Starting cleanup of old public data" celery.log
+```
+
+Task statistics are logged after each run:
+- Number of transcripts deleted
+- Number of meetings deleted
+- Number of orphaned recordings deleted
+- Any errors encountered

server/docs/gpu/api-transcription.md

@@ -0,0 +1,194 @@
+## Reflector GPU Transcription API (Specification)
+
+This document defines the Reflector GPU transcription API that all implementations must adhere to. Current implementations include NVIDIA Parakeet (NeMo) and Whisper (faster-whisper), both deployed on Modal.com. The API surface and response shapes are OpenAI/Whisper-compatible, so clients can switch implementations by changing only the base URL.
+
+### Base URL and Authentication
+
+- Example base URLs (Modal web endpoints):
+
+  - Parakeet: `https://<account>--reflector-transcriber-parakeet-web.modal.run`
+  - Whisper: `https://<account>--reflector-transcriber-web.modal.run`
+
+- All endpoints are served under `/v1` and require a Bearer token:
+
+```
+Authorization: Bearer <REFLECTOR_GPU_APIKEY>
+```
+
+Note: To switch implementations, deploy the desired variant and point `TRANSCRIPT_URL` to its base URL. The API is identical.
+
+### Supported file types
+
+`mp3, mp4, mpeg, mpga, m4a, wav, webm`
+
+### Models and languages
+
+- Parakeet (NVIDIA NeMo): default `nvidia/parakeet-tdt-0.6b-v2`
+  - Language support: only `en`. Other languages return HTTP 400.
+- Whisper (faster-whisper): default `large-v2` (or deployment-specific)
+  - Language support: multilingual (per Whisper model capabilities).
+
+Note: The `model` parameter is accepted by all implementations for interface parity. Some backends may treat it as informational.
+
+### Endpoints
+
+#### POST /v1/audio/transcriptions
+
+Transcribe one or more uploaded audio files.
+
+Request: multipart/form-data
+
+- `file` (File) — optional. Single file to transcribe.
+- `files` (File[]) — optional. One or more files to transcribe.
+- `model` (string) — optional. Defaults to the implementation-specific model (see above).
+- `language` (string) — optional, defaults to `en`.
+  - Parakeet: only `en` is accepted; other values return HTTP 400
+  - Whisper: model-dependent; typically multilingual
+- `batch` (boolean) — optional, defaults to `false`.
+
+Notes:
+
+- Provide either `file` or `files`, not both. If neither is provided, HTTP 400.
+- `batch` requires `files`; using `batch=true` without `files` returns HTTP 400.
+- Response shape for multiple files is the same regardless of `batch`.
+- Files sent to this endpoint are processed in a single pass (no VAD/chunking). This is intended for short clips (roughly ≤ 30s; depends on GPU memory/model). For longer audio, prefer `/v1/audio/transcriptions-from-url` which supports VAD-based chunking.
+
+Responses
+
+Single file response:
+
+```json
+{
+  "text": "transcribed text",
+  "words": [
+    { "word": "hello", "start": 0.0, "end": 0.5 },
+    { "word": "world", "start": 0.5, "end": 1.0 }
+  ],
+  "filename": "audio.mp3"
+}
+```
+
+Multiple files response:
+
+```json
+{
+  "results": [
+    {"filename": "a1.mp3", "text": "...", "words": [...]},
+    {"filename": "a2.mp3", "text": "...", "words": [...]}]
+}
+```
+
+Notes:
+
+- Word objects always include keys: `word`, `start`, `end`.
+- Some implementations may include a trailing space in `word` to match Whisper tokenization behavior; clients should trim if needed.
+
+Example curl (single file):
+
+```bash
+curl -X POST \
+  -H "Authorization: Bearer $REFLECTOR_GPU_APIKEY" \
+  -F "file=@/path/to/audio.mp3" \
+  -F "language=en" \
+  "$BASE_URL/v1/audio/transcriptions"
+```
+
+Example curl (multiple files, batch):
+
+```bash
+curl -X POST \
+  -H "Authorization: Bearer $REFLECTOR_GPU_APIKEY" \
+  -F "files=@/path/a1.mp3" -F "files=@/path/a2.mp3" \
+  -F "batch=true" -F "language=en" \
+  "$BASE_URL/v1/audio/transcriptions"
+```
+
+#### POST /v1/audio/transcriptions-from-url
+
+Transcribe a single remote audio file by URL.
+
+Request: application/json
+
+Body parameters:
+
+- `audio_file_url` (string) — required. URL of the audio file to transcribe.
+- `model` (string) — optional. Defaults to the implementation-specific model (see above).
+- `language` (string) — optional, defaults to `en`. Parakeet only accepts `en`.
+- `timestamp_offset` (number) — optional, defaults to `0.0`. Added to each word's `start`/`end` in the response.
+
+```json
+{
+  "audio_file_url": "https://example.com/audio.mp3",
+  "model": "nvidia/parakeet-tdt-0.6b-v2",
+  "language": "en",
+  "timestamp_offset": 0.0
+}
+```
+
+Response:
+
+```json
+{
+  "text": "transcribed text",
+  "words": [
+    { "word": "hello", "start": 10.0, "end": 10.5 },
+    { "word": "world", "start": 10.5, "end": 11.0 }
+  ]
+}
+```
+
+Notes:
+
+- `timestamp_offset` is added to each word’s `start`/`end` in the response.
+- Implementations may perform VAD-based chunking and batching for long-form audio; word timings are adjusted accordingly.
+
+Example curl:
+
+```bash
+curl -X POST \
+  -H "Authorization: Bearer $REFLECTOR_GPU_APIKEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+        "audio_file_url": "https://example.com/audio.mp3",
+        "language": "en",
+        "timestamp_offset": 0
+      }' \
+  "$BASE_URL/v1/audio/transcriptions-from-url"
+```
+
+### Error handling
+
+- 400 Bad Request
+  - Parakeet: `language` other than `en`
+  - Missing required parameters (`file`/`files` for upload; `audio_file_url` for URL endpoint)
+  - Unsupported file extension
+- 401 Unauthorized
+  - Missing or invalid Bearer token
+- 404 Not Found
+  - `audio_file_url` does not exist
+
+### Implementation details
+
+- GPUs: A10G for small-file/live, L40S for large-file URL transcription (subject to deployment)
+- VAD chunking and segment batching; word timings adjusted and overlapping ends constrained
+- Pads very short segments (< 0.5s) to avoid model crashes on some backends
+
+### Server configuration (Reflector API)
+
+Set the Reflector server to use the Modal backend and point `TRANSCRIPT_URL` to your chosen deployment:
+
+```
+TRANSCRIPT_BACKEND=modal
+TRANSCRIPT_URL=https://<account>--reflector-transcriber-parakeet-web.modal.run
+TRANSCRIPT_MODAL_API_KEY=<REFLECTOR_GPU_APIKEY>
+```
+
+### Conformance tests
+
+Use the pytest-based conformance tests to validate any new implementation (including self-hosted) against this spec:
+
+```
+TRANSCRIPT_URL=https://<your-deployment-base> \
+TRANSCRIPT_MODAL_API_KEY=your-api-key \
+uv run -m pytest -m model_api --no-cov server/tests/test_model_api_transcript.py
+```

server/docs/webhook.md

@@ -0,0 +1,233 @@
+# Reflector Webhook Documentation
+
+## Overview
+
+Reflector supports webhook notifications to notify external systems when transcript processing is completed. Webhooks can be configured per room and are triggered automatically after a transcript is successfully processed.
+
+## Configuration
+
+Webhooks are configured at the room level with two fields:
+- `webhook_url`: The HTTPS endpoint to receive webhook notifications
+- `webhook_secret`: Optional secret key for HMAC signature verification (auto-generated if not provided)
+
+## Events
+
+### `transcript.completed`
+
+Triggered when a transcript has been fully processed, including transcription, diarization, summarization, topic detection and calendar event integration.
+
+### `test`
+
+A test event that can be triggered manually to verify webhook configuration.
+
+## Webhook Request Format
+
+### Headers
+
+All webhook requests include the following headers:
+
+| Header | Description | Example |
+|--------|-------------|---------|
+| `Content-Type` | Always `application/json` | `application/json` |
+| `User-Agent` | Identifies Reflector as the source | `Reflector-Webhook/1.0` |
+| `X-Webhook-Event` | The event type | `transcript.completed` or `test` |
+| `X-Webhook-Retry` | Current retry attempt number | `0`, `1`, `2`... |
+| `X-Webhook-Signature` | HMAC signature (if secret configured) | `t=1735306800,v1=abc123...` |
+
+### Signature Verification
+
+If a webhook secret is configured, Reflector includes an HMAC-SHA256 signature in the `X-Webhook-Signature` header to verify the webhook authenticity.
+
+The signature format is: `t={timestamp},v1={signature}`
+
+To verify the signature:
+1. Extract the timestamp and signature from the header
+2. Create the signed payload: `{timestamp}.{request_body}`
+3. Compute HMAC-SHA256 of the signed payload using your webhook secret
+4. Compare the computed signature with the received signature
+
+Example verification (Python):
+```python
+import hmac
+import hashlib
+
+def verify_webhook_signature(payload: bytes, signature_header: str, secret: str) -> bool:
+    # Parse header: "t=1735306800,v1=abc123..."
+    parts = dict(part.split("=") for part in signature_header.split(","))
+    timestamp = parts["t"]
+    received_signature = parts["v1"]
+
+    # Create signed payload
+    signed_payload = f"{timestamp}.{payload.decode('utf-8')}"
+
+    # Compute expected signature
+    expected_signature = hmac.new(
+        secret.encode("utf-8"),
+        signed_payload.encode("utf-8"),
+        hashlib.sha256
+    ).hexdigest()
+
+    # Compare signatures
+    return hmac.compare_digest(expected_signature, received_signature)
+```
+
+## Event Payloads
+
+### `transcript.completed` Event
+
+This event includes a convenient URL for accessing the transcript:
+- `frontend_url`: Direct link to view the transcript in the web interface
+
+```json
+{
+  "event": "transcript.completed",
+  "event_id": "transcript.completed-abc-123-def-456",
+  "timestamp": "2025-08-27T12:34:56.789012Z",
+  "transcript": {
+    "id": "abc-123-def-456",
+    "room_id": "room-789",
+    "created_at": "2025-08-27T12:00:00Z",
+    "duration": 1800.5,
+    "title": "Q3 Product Planning Meeting",
+    "short_summary": "Team discussed Q3 product roadmap, prioritizing mobile app features and API improvements.",
+    "long_summary": "The product team met to finalize the Q3 roadmap. Key decisions included...",
+    "webvtt": "WEBVTT\n\n00:00:00.000 --> 00:00:05.000\n<v Speaker 1>Welcome everyone to today's meeting...",
+    "topics": [
+      {
+        "title": "Introduction and Agenda",
+        "summary": "Meeting kickoff with agenda review",
+        "timestamp": 0.0,
+        "duration": 120.0,
+        "webvtt": "WEBVTT\n\n00:00:00.000 --> 00:00:05.000\n<v Speaker 1>Welcome everyone..."
+      },
+      {
+        "title": "Mobile App Features Discussion",
+        "summary": "Team reviewed proposed mobile app features for Q3",
+        "timestamp": 120.0,
+        "duration": 600.0,
+        "webvtt": "WEBVTT\n\n00:02:00.000 --> 00:02:10.000\n<v Speaker 2>Let's talk about the mobile app..."
+      }
+    ],
+    "participants": [
+      {
+        "id": "participant-1",
+        "name": "John Doe",
+        "speaker": "Speaker 1"
+      },
+      {
+        "id": "participant-2",
+        "name": "Jane Smith",
+        "speaker": "Speaker 2"
+      }
+    ],
+    "source_language": "en",
+    "target_language": "en",
+    "status": "completed",
+    "frontend_url": "https://app.reflector.com/transcripts/abc-123-def-456"
+  },
+  "room": {
+    "id": "room-789",
+    "name": "Product Team Room"
+  },
+  "calendar_event": {
+    "id": "calendar-event-123",
+    "ics_uid": "event-123",
+    "title": "Q3 Product Planning Meeting",
+    "start_time": "2025-08-27T12:00:00Z",
+    "end_time": "2025-08-27T12:30:00Z",
+    "description": "Team discussed Q3 product roadmap, prioritizing mobile app features and API improvements.",
+    "location": "Conference Room 1",
+    "attendees": [
+      {
+        "id": "participant-1",
+        "name": "John Doe",
+        "speaker": "Speaker 1"
+      },
+      {
+        "id": "participant-2",
+        "name": "Jane Smith",
+        "speaker": "Speaker 2"
+      }
+    ]
+  }
+}
+```
+
+### `test` Event
+
+```json
+{
+  "event": "test",
+  "event_id": "test.2025-08-27T12:34:56.789012Z",
+  "timestamp": "2025-08-27T12:34:56.789012Z",
+  "message": "This is a test webhook from Reflector",
+  "room": {
+    "id": "room-789",
+    "name": "Product Team Room"
+  }
+}
+```
+
+## Retry Policy
+
+Webhooks are delivered with automatic retry logic to handle transient failures. When a webhook delivery fails due to server errors or network issues, Reflector will automatically retry the delivery multiple times over an extended period.
+
+### Retry Mechanism
+
+Reflector implements an exponential backoff strategy for webhook retries:
+
+- **Initial retry delay**: 60 seconds after the first failure
+- **Exponential backoff**: Each subsequent retry waits approximately twice as long as the previous one
+- **Maximum retry interval**: 1 hour (backoff is capped at this duration)
+- **Maximum retry attempts**: 30 attempts total
+- **Total retry duration**: Retries continue for approximately 24 hours
+
+### How Retries Work
+
+When a webhook fails, Reflector will:
+1. Wait 60 seconds, then retry (attempt #1)
+2. If it fails again, wait ~2 minutes, then retry (attempt #2)
+3. Continue doubling the wait time up to a maximum of 1 hour between attempts
+4. Keep retrying at 1-hour intervals until successful or 30 attempts are exhausted
+
+The `X-Webhook-Retry` header indicates the current retry attempt number (0 for the initial attempt, 1 for first retry, etc.), allowing your endpoint to track retry attempts.
+
+### Retry Behavior by HTTP Status Code
+
+| Status Code | Behavior |
+|-------------|----------|
+| 2xx (Success) | No retry, webhook marked as delivered |
+| 4xx (Client Error) | No retry, request is considered permanently failed |
+| 5xx (Server Error) | Automatic retry with exponential backoff |
+| Network/Timeout Error | Automatic retry with exponential backoff |
+
+**Important Notes:**
+- Webhooks timeout after 30 seconds. If your endpoint takes longer to respond, it will be considered a timeout error and retried.
+- During the retry period (~24 hours), you may receive the same webhook multiple times if your endpoint experiences intermittent failures.
+- There is no mechanism to manually retry failed webhooks after the retry period expires.
+
+## Testing Webhooks
+
+You can test your webhook configuration before processing transcripts:
+
+```http
+POST /v1/rooms/{room_id}/webhook/test
+```
+
+Response:
+```json
+{
+  "success": true,
+  "status_code": 200,
+  "message": "Webhook test successful",
+  "response_preview": "OK"
+}
+```
+
+Or in case of failure:
+```json
+{
+  "success": false,
+  "error": "Webhook request timed out (10 seconds)"
+}
+```

server/env.example

@@ -27,7 +27,7 @@ AUTH_JWT_AUDIENCE=
 #TRANSCRIPT_MODAL_API_KEY=xxxxx
 
 TRANSCRIPT_BACKEND=modal
-TRANSCRIPT_URL=https://monadical-sas--reflector-transcriber-web.modal.run
+TRANSCRIPT_URL=https://monadical-sas--reflector-transcriber-parakeet-web.modal.run
 TRANSCRIPT_MODAL_API_KEY=
 
 ## =======================================================
@@ -71,3 +71,27 @@ DIARIZATION_URL=https://monadical-sas--reflector-diarizer-web.modal.run
 
 ## Sentry DSN configuration
 #SENTRY_DSN=
+
+## =======================================================
+## Video Platform Configuration
+## =======================================================
+
+## Whereby
+#WHEREBY_API_KEY=your-whereby-api-key
+#WHEREBY_WEBHOOK_SECRET=your-whereby-webhook-secret
+#AWS_WHEREBY_ACCESS_KEY_ID=your-aws-key
+#AWS_WHEREBY_ACCESS_KEY_SECRET=your-aws-secret
+#AWS_PROCESS_RECORDING_QUEUE_URL=https://sqs.us-west-2.amazonaws.com/...
+
+## Daily.co
+#DAILY_API_KEY=your-daily-api-key
+#DAILY_WEBHOOK_SECRET=your-daily-webhook-secret
+#DAILY_SUBDOMAIN=your-subdomain
+#AWS_DAILY_S3_BUCKET=your-daily-bucket
+#AWS_DAILY_S3_REGION=us-west-2
+#AWS_DAILY_ROLE_ARN=arn:aws:iam::ACCOUNT:role/DailyRecording
+
+## Platform Selection
+#DAILY_MIGRATION_ENABLED=false           # Enable Daily.co support
+#DAILY_MIGRATION_ROOM_IDS=[]            # Specific rooms to use Daily
+#DEFAULT_VIDEO_PLATFORM=whereby          # Default platform for new rooms

server/gpu/modal_deployments/reflector_transcriber.py

@@ -1,161 +0,0 @@
-import os
-import tempfile
-import threading
-
-import modal
-from pydantic import BaseModel
-
-MODELS_DIR = "/models"
-
-MODEL_NAME = "large-v2"
-MODEL_COMPUTE_TYPE: str = "float16"
-MODEL_NUM_WORKERS: int = 1
-
-MINUTES = 60  # seconds
-
-volume = modal.Volume.from_name("models", create_if_missing=True)
-
-app = modal.App("reflector-transcriber")
-
-
-def download_model():
-    from faster_whisper import download_model
-
-    volume.reload()
-
-    download_model(MODEL_NAME, cache_dir=MODELS_DIR)
-
-    volume.commit()
-
-
-image = (
-    modal.Image.debian_slim(python_version="3.12")
-    .pip_install(
-        "huggingface_hub==0.27.1",
-        "hf-transfer==0.1.9",
-        "torch==2.5.1",
-        "faster-whisper==1.1.1",
-    )
-    .env(
-        {
-            "HF_HUB_ENABLE_HF_TRANSFER": "1",
-            "LD_LIBRARY_PATH": (
-                "/usr/local/lib/python3.12/site-packages/nvidia/cudnn/lib/:"
-                "/opt/conda/lib/python3.12/site-packages/nvidia/cublas/lib/"
-            ),
-        }
-    )
-    .run_function(download_model, volumes={MODELS_DIR: volume})
-)
-
-
-@app.cls(
-    gpu="A10G",
-    timeout=5 * MINUTES,
-    scaledown_window=5 * MINUTES,
-    allow_concurrent_inputs=6,
-    image=image,
-    volumes={MODELS_DIR: volume},
-)
-class Transcriber:
-    @modal.enter()
-    def enter(self):
-        import faster_whisper
-        import torch
-
-        self.lock = threading.Lock()
-        self.use_gpu = torch.cuda.is_available()
-        self.device = "cuda" if self.use_gpu else "cpu"
-        self.model = faster_whisper.WhisperModel(
-            MODEL_NAME,
-            device=self.device,
-            compute_type=MODEL_COMPUTE_TYPE,
-            num_workers=MODEL_NUM_WORKERS,
-            download_root=MODELS_DIR,
-            local_files_only=True,
-        )
-
-    @modal.method()
-    def transcribe_segment(
-        self,
-        audio_data: str,
-        audio_suffix: str,
-        language: str,
-    ):
-        with tempfile.NamedTemporaryFile("wb+", suffix=f".{audio_suffix}") as fp:
-            fp.write(audio_data)
-
-            with self.lock:
-                segments, _ = self.model.transcribe(
-                    fp.name,
-                    language=language,
-                    beam_size=5,
-                    word_timestamps=True,
-                    vad_filter=True,
-                    vad_parameters={"min_silence_duration_ms": 500},
-                )
-
-            segments = list(segments)
-            text = "".join(segment.text for segment in segments)
-            words = [
-                {"word": word.word, "start": word.start, "end": word.end}
-                for segment in segments
-                for word in segment.words
-            ]
-
-            return {"text": text, "words": words}
-
-
-@app.function(
-    scaledown_window=60,
-    timeout=60,
-    allow_concurrent_inputs=40,
-    secrets=[
-        modal.Secret.from_name("reflector-gpu"),
-    ],
-    volumes={MODELS_DIR: volume},
-)
-@modal.asgi_app()
-def web():
-    from fastapi import Body, Depends, FastAPI, HTTPException, UploadFile, status
-    from fastapi.security import OAuth2PasswordBearer
-    from typing_extensions import Annotated
-
-    transcriber = Transcriber()
-
-    app = FastAPI()
-
-    oauth2_scheme = OAuth2PasswordBearer(tokenUrl="token")
-
-    supported_file_types = ["mp3", "mp4", "mpeg", "mpga", "m4a", "wav", "webm"]
-
-    def apikey_auth(apikey: str = Depends(oauth2_scheme)):
-        if apikey != os.environ["REFLECTOR_GPU_APIKEY"]:
-            raise HTTPException(
-                status_code=status.HTTP_401_UNAUTHORIZED,
-                detail="Invalid API key",
-                headers={"WWW-Authenticate": "Bearer"},
-            )
-
-    class TranscriptResponse(BaseModel):
-        result: dict
-
-    @app.post("/v1/audio/transcriptions", dependencies=[Depends(apikey_auth)])
-    def transcribe(
-        file: UploadFile,
-        model: str = "whisper-1",
-        language: Annotated[str, Body(...)] = "en",
-    ) -> TranscriptResponse:
-        audio_data = file.file.read()
-        audio_suffix = file.filename.split(".")[-1]
-        assert audio_suffix in supported_file_types
-
-        func = transcriber.transcribe_segment.spawn(
-            audio_data=audio_data,
-            audio_suffix=audio_suffix,
-            language=language,
-        )
-        result = func.get()
-        return result
-
-    return app

server/gpu/modal_deployments/reflector_transcriber_parakeet_v3.py

@@ -1,622 +0,0 @@
-import logging
-import os
-import sys
-import threading
-import uuid
-from typing import Mapping, NewType
-from urllib.parse import urlparse
-
-import modal
-
-MODEL_NAME = "nvidia/parakeet-tdt-0.6b-v3"
-SUPPORTED_FILE_EXTENSIONS = ["mp3", "mp4", "mpeg", "mpga", "m4a", "wav", "webm"]
-SAMPLERATE = 16000
-UPLOADS_PATH = "/uploads"
-CACHE_PATH = "/cache"
-VAD_CONFIG = {
-    "max_segment_duration": 30.0,
-    "batch_max_files": 10,
-    "batch_max_duration": 5.0,
-    "min_segment_duration": 0.02,
-    "silence_padding": 0.5,
-    "window_size": 512,
-}
-
-ParakeetUniqFilename = NewType("ParakeetUniqFilename", str)
-AudioFileExtension = NewType("AudioFileExtension", str)
-
-app = modal.App("reflector-transcriber-parakeet-v3")
-
-# Volume for caching model weights
-model_cache = modal.Volume.from_name("parakeet-model-cache", create_if_missing=True)
-# Volume for temporary file uploads
-upload_volume = modal.Volume.from_name("parakeet-uploads", create_if_missing=True)
-
-image = (
-    modal.Image.from_registry(
-        "nvidia/cuda:12.8.0-cudnn-devel-ubuntu22.04", add_python="3.12"
-    )
-    .env(
-        {
-            "HF_HUB_ENABLE_HF_TRANSFER": "1",
-            "HF_HOME": "/cache",
-            "DEBIAN_FRONTEND": "noninteractive",
-            "CXX": "g++",
-            "CC": "g++",
-        }
-    )
-    .apt_install("ffmpeg")
-    .pip_install(
-        "hf_transfer==0.1.9",
-        "huggingface_hub[hf-xet]==0.31.2",
-        "nemo_toolkit[asr]==2.3.0",
-        "cuda-python==12.8.0",
-        "fastapi==0.115.12",
-        "numpy<2",
-        "librosa==0.10.1",
-        "requests",
-        "silero-vad==5.1.0",
-        "torch",
-    )
-    .entrypoint([])  # silence chatty logs by container on start
-)
-
-
-def detect_audio_format(url: str, headers: Mapping[str, str]) -> AudioFileExtension:
-    parsed_url = urlparse(url)
-    url_path = parsed_url.path
-
-    for ext in SUPPORTED_FILE_EXTENSIONS:
-        if url_path.lower().endswith(f".{ext}"):
-            return AudioFileExtension(ext)
-
-    content_type = headers.get("content-type", "").lower()
-    if "audio/mpeg" in content_type or "audio/mp3" in content_type:
-        return AudioFileExtension("mp3")
-    if "audio/wav" in content_type:
-        return AudioFileExtension("wav")
-    if "audio/mp4" in content_type:
-        return AudioFileExtension("mp4")
-
-    raise ValueError(
-        f"Unsupported audio format for URL: {url}. "
-        f"Supported extensions: {', '.join(SUPPORTED_FILE_EXTENSIONS)}"
-    )
-
-
-def download_audio_to_volume(
-    audio_file_url: str,
-) -> tuple[ParakeetUniqFilename, AudioFileExtension]:
-    import requests
-    from fastapi import HTTPException
-
-    response = requests.head(audio_file_url, allow_redirects=True)
-    if response.status_code == 404:
-        raise HTTPException(status_code=404, detail="Audio file not found")
-
-    response = requests.get(audio_file_url, allow_redirects=True)
-    response.raise_for_status()
-
-    audio_suffix = detect_audio_format(audio_file_url, response.headers)
-    unique_filename = ParakeetUniqFilename(f"{uuid.uuid4()}.{audio_suffix}")
-    file_path = f"{UPLOADS_PATH}/{unique_filename}"
-
-    with open(file_path, "wb") as f:
-        f.write(response.content)
-
-    upload_volume.commit()
-    return unique_filename, audio_suffix
-
-
-def pad_audio(audio_array, sample_rate: int = SAMPLERATE):
-    """Add 0.5 seconds of silence if audio is less than 500ms.
-
-    This is a workaround for a Parakeet bug where very short audio (<500ms) causes:
-    ValueError: `char_offsets`: [] and `processed_tokens`: [157, 834, 834, 841]
-    have to be of the same length
-
-    See: https://github.com/NVIDIA/NeMo/issues/8451
-    """
-    import numpy as np
-
-    audio_duration = len(audio_array) / sample_rate
-    if audio_duration < 0.5:
-        silence_samples = int(sample_rate * 0.5)
-        silence = np.zeros(silence_samples, dtype=np.float32)
-        return np.concatenate([audio_array, silence])
-    return audio_array
-
-
-@app.cls(
-    gpu="A10G",
-    timeout=600,
-    scaledown_window=300,
-    image=image,
-    volumes={CACHE_PATH: model_cache, UPLOADS_PATH: upload_volume},
-    enable_memory_snapshot=True,
-    experimental_options={"enable_gpu_snapshot": True},
-)
-@modal.concurrent(max_inputs=10)
-class TranscriberParakeetLive:
-    @modal.enter(snap=True)
-    def enter(self):
-        import nemo.collections.asr as nemo_asr
-
-        logging.getLogger("nemo_logger").setLevel(logging.CRITICAL)
-
-        self.lock = threading.Lock()
-        self.model = nemo_asr.models.ASRModel.from_pretrained(model_name=MODEL_NAME)
-        device = next(self.model.parameters()).device
-        print(f"Model is on device: {device}")
-
-    @modal.method()
-    def transcribe_segment(
-        self,
-        filename: str,
-    ):
-        import librosa
-
-        upload_volume.reload()
-
-        file_path = f"{UPLOADS_PATH}/{filename}"
-        if not os.path.exists(file_path):
-            raise FileNotFoundError(f"File not found: {file_path}")
-
-        audio_array, sample_rate = librosa.load(file_path, sr=SAMPLERATE, mono=True)
-        padded_audio = pad_audio(audio_array, sample_rate)
-
-        with self.lock:
-            with NoStdStreams():
-                (output,) = self.model.transcribe([padded_audio], timestamps=True)
-
-        text = output.text.strip()
-        words = [
-            {
-                "word": word_info["word"] + " ",
-                "start": round(word_info["start"], 2),
-                "end": round(word_info["end"], 2),
-            }
-            for word_info in output.timestamp["word"]
-        ]
-
-        return {"text": text, "words": words}
-
-    @modal.method()
-    def transcribe_batch(
-        self,
-        filenames: list[str],
-    ):
-        import librosa
-
-        upload_volume.reload()
-
-        results = []
-        audio_arrays = []
-
-        # Load all audio files with padding
-        for filename in filenames:
-            file_path = f"{UPLOADS_PATH}/{filename}"
-            if not os.path.exists(file_path):
-                raise FileNotFoundError(f"Batch file not found: {file_path}")
-
-            audio_array, sample_rate = librosa.load(file_path, sr=SAMPLERATE, mono=True)
-            padded_audio = pad_audio(audio_array, sample_rate)
-            audio_arrays.append(padded_audio)
-
-        with self.lock:
-            with NoStdStreams():
-                outputs = self.model.transcribe(audio_arrays, timestamps=True)
-
-        # Process results for each file
-        for i, (filename, output) in enumerate(zip(filenames, outputs)):
-            text = output.text.strip()
-
-            words = [
-                {
-                    "word": word_info["word"] + " ",
-                    "start": round(word_info["start"], 2),
-                    "end": round(word_info["end"], 2),
-                }
-                for word_info in output.timestamp["word"]
-            ]
-
-            results.append(
-                {
-                    "filename": filename,
-                    "text": text,
-                    "words": words,
-                }
-            )
-
-        return results
-
-
-# L40S class for file transcription (bigger files)
-@app.cls(
-    gpu="L40S",
-    timeout=900,
-    image=image,
-    volumes={CACHE_PATH: model_cache, UPLOADS_PATH: upload_volume},
-    enable_memory_snapshot=True,
-    experimental_options={"enable_gpu_snapshot": True},
-)
-class TranscriberParakeetFile:
-    @modal.enter(snap=True)
-    def enter(self):
-        import nemo.collections.asr as nemo_asr
-        import torch
-        from silero_vad import load_silero_vad
-
-        logging.getLogger("nemo_logger").setLevel(logging.CRITICAL)
-
-        self.model = nemo_asr.models.ASRModel.from_pretrained(model_name=MODEL_NAME)
-        device = next(self.model.parameters()).device
-        print(f"Model is on device: {device}")
-
-        torch.set_num_threads(1)
-        self.vad_model = load_silero_vad(onnx=False)
-        print("Silero VAD initialized")
-
-    @modal.method()
-    def transcribe_segment(
-        self,
-        filename: str,
-        timestamp_offset: float = 0.0,
-    ):
-        import librosa
-        import numpy as np
-        from silero_vad import VADIterator
-
-        def load_and_convert_audio(file_path):
-            audio_array, sample_rate = librosa.load(file_path, sr=SAMPLERATE, mono=True)
-            return audio_array
-
-        def vad_segment_generator(audio_array):
-            """Generate speech segments using VAD with start/end sample indices"""
-            vad_iterator = VADIterator(self.vad_model, sampling_rate=SAMPLERATE)
-            window_size = VAD_CONFIG["window_size"]
-            start = None
-
-            for i in range(0, len(audio_array), window_size):
-                chunk = audio_array[i : i + window_size]
-                if len(chunk) < window_size:
-                    chunk = np.pad(
-                        chunk, (0, window_size - len(chunk)), mode="constant"
-                    )
-
-                speech_dict = vad_iterator(chunk)
-                if not speech_dict:
-                    continue
-
-                if "start" in speech_dict:
-                    start = speech_dict["start"]
-                    continue
-
-                if "end" in speech_dict and start is not None:
-                    end = speech_dict["end"]
-                    start_time = start / float(SAMPLERATE)
-                    end_time = end / float(SAMPLERATE)
-
-                    # Extract the actual audio segment
-                    audio_segment = audio_array[start:end]
-
-                    yield (start_time, end_time, audio_segment)
-                    start = None
-
-            vad_iterator.reset_states()
-
-        def vad_segment_filter(segments):
-            """Filter VAD segments by duration and chunk large segments"""
-            min_dur = VAD_CONFIG["min_segment_duration"]
-            max_dur = VAD_CONFIG["max_segment_duration"]
-
-            for start_time, end_time, audio_segment in segments:
-                segment_duration = end_time - start_time
-
-                # Skip very small segments
-                if segment_duration < min_dur:
-                    continue
-
-                # If segment is within max duration, yield as-is
-                if segment_duration <= max_dur:
-                    yield (start_time, end_time, audio_segment)
-                    continue
-
-                # Chunk large segments into smaller pieces
-                chunk_samples = int(max_dur * SAMPLERATE)
-                current_start = start_time
-
-                for chunk_offset in range(0, len(audio_segment), chunk_samples):
-                    chunk_audio = audio_segment[
-                        chunk_offset : chunk_offset + chunk_samples
-                    ]
-                    if len(chunk_audio) == 0:
-                        break
-
-                    chunk_duration = len(chunk_audio) / float(SAMPLERATE)
-                    chunk_end = current_start + chunk_duration
-
-                    # Only yield chunks that meet minimum duration
-                    if chunk_duration >= min_dur:
-                        yield (current_start, chunk_end, chunk_audio)
-
-                    current_start = chunk_end
-
-        def batch_segments(segments, max_files=10, max_duration=5.0):
-            batch = []
-            batch_duration = 0.0
-
-            for start_time, end_time, audio_segment in segments:
-                segment_duration = end_time - start_time
-
-                if segment_duration < VAD_CONFIG["silence_padding"]:
-                    silence_samples = int(
-                        (VAD_CONFIG["silence_padding"] - segment_duration) * SAMPLERATE
-                    )
-                    padding = np.zeros(silence_samples, dtype=np.float32)
-                    audio_segment = np.concatenate([audio_segment, padding])
-                    segment_duration = VAD_CONFIG["silence_padding"]
-
-                batch.append((start_time, end_time, audio_segment))
-                batch_duration += segment_duration
-
-                if len(batch) >= max_files or batch_duration >= max_duration:
-                    yield batch
-                    batch = []
-                    batch_duration = 0.0
-
-            if batch:
-                yield batch
-
-        def transcribe_batch(model, audio_segments):
-            with NoStdStreams():
-                outputs = model.transcribe(audio_segments, timestamps=True)
-            return outputs
-
-        def emit_results(
-            results,
-            segments_info,
-            batch_index,
-            total_batches,
-        ):
-            """Yield transcribed text and word timings from model output, adjusting timestamps to absolute positions."""
-            for i, (output, (start_time, end_time, _)) in enumerate(
-                zip(results, segments_info)
-            ):
-                text = output.text.strip()
-                words = [
-                    {
-                        "word": word_info["word"],
-                        "start": round(
-                            word_info["start"] + start_time + timestamp_offset, 2
-                        ),
-                        "end": round(
-                            word_info["end"] + start_time + timestamp_offset, 2
-                        ),
-                    }
-                    for word_info in output.timestamp["word"]
-                ]
-
-                yield text, words
-
-        upload_volume.reload()
-
-        file_path = f"{UPLOADS_PATH}/{filename}"
-        if not os.path.exists(file_path):
-            raise FileNotFoundError(f"File not found: {file_path}")
-
-        audio_array = load_and_convert_audio(file_path)
-        total_duration = len(audio_array) / float(SAMPLERATE)
-        processed_duration = 0.0
-
-        all_text_parts = []
-        all_words = []
-
-        raw_segments = vad_segment_generator(audio_array)
-        filtered_segments = vad_segment_filter(raw_segments)
-        batches = batch_segments(
-            filtered_segments,
-            VAD_CONFIG["batch_max_files"],
-            VAD_CONFIG["batch_max_duration"],
-        )
-
-        batch_index = 0
-        total_batches = max(
-            1, int(total_duration / VAD_CONFIG["batch_max_duration"]) + 1
-        )
-
-        for batch in batches:
-            batch_index += 1
-            audio_segments = [seg[2] for seg in batch]
-            results = transcribe_batch(self.model, audio_segments)
-
-            for text, words in emit_results(
-                results,
-                batch,
-                batch_index,
-                total_batches,
-            ):
-                if not text:
-                    continue
-                all_text_parts.append(text)
-                all_words.extend(words)
-
-            processed_duration += sum(len(seg[2]) / float(SAMPLERATE) for seg in batch)
-
-        combined_text = " ".join(all_text_parts)
-        return {"text": combined_text, "words": all_words}
-
-
-@app.function(
-    scaledown_window=60,
-    timeout=600,
-    secrets=[
-        modal.Secret.from_name("reflector-gpu"),
-    ],
-    volumes={CACHE_PATH: model_cache, UPLOADS_PATH: upload_volume},
-    image=image,
-)
-@modal.concurrent(max_inputs=40)
-@modal.asgi_app()
-def web():
-    import os
-    import uuid
-
-    from fastapi import (
-        Body,
-        Depends,
-        FastAPI,
-        Form,
-        HTTPException,
-        UploadFile,
-        status,
-    )
-    from fastapi.security import OAuth2PasswordBearer
-    from pydantic import BaseModel
-
-    transcriber_live = TranscriberParakeetLive()
-    transcriber_file = TranscriberParakeetFile()
-
-    app = FastAPI()
-
-    oauth2_scheme = OAuth2PasswordBearer(tokenUrl="token")
-
-    def apikey_auth(apikey: str = Depends(oauth2_scheme)):
-        if apikey == os.environ["REFLECTOR_GPU_APIKEY"]:
-            return
-        raise HTTPException(
-            status_code=status.HTTP_401_UNAUTHORIZED,
-            detail="Invalid API key",
-            headers={"WWW-Authenticate": "Bearer"},
-        )
-
-    class TranscriptResponse(BaseModel):
-        result: dict
-
-    @app.post("/v1/audio/transcriptions", dependencies=[Depends(apikey_auth)])
-    def transcribe(
-        file: UploadFile = None,
-        files: list[UploadFile] | None = None,
-        model: str = Form(MODEL_NAME),
-        language: str = Form("en"),
-        batch: bool = Form(False),
-    ):
-        # Parakeet only supports English
-        if language != "en":
-            raise HTTPException(
-                status_code=400,
-                detail=f"Parakeet model only supports English. Got language='{language}'",
-            )
-        # Handle both single file and multiple files
-        if not file and not files:
-            raise HTTPException(
-                status_code=400, detail="Either 'file' or 'files' parameter is required"
-            )
-        if batch and not files:
-            raise HTTPException(
-                status_code=400, detail="Batch transcription requires 'files'"
-            )
-
-        upload_files = [file] if file else files
-
-        # Upload files to volume
-        uploaded_filenames = []
-        for upload_file in upload_files:
-            audio_suffix = upload_file.filename.split(".")[-1]
-            assert audio_suffix in SUPPORTED_FILE_EXTENSIONS
-
-            # Generate unique filename
-            unique_filename = f"{uuid.uuid4()}.{audio_suffix}"
-            file_path = f"{UPLOADS_PATH}/{unique_filename}"
-
-            print(f"Writing file to: {file_path}")
-            with open(file_path, "wb") as f:
-                content = upload_file.file.read()
-                f.write(content)
-
-            uploaded_filenames.append(unique_filename)
-
-        upload_volume.commit()
-
-        try:
-            # Use A10G live transcriber for per-file transcription
-            if batch and len(upload_files) > 1:
-                # Use batch transcription
-                func = transcriber_live.transcribe_batch.spawn(
-                    filenames=uploaded_filenames,
-                )
-                results = func.get()
-                return {"results": results}
-
-            # Per-file transcription
-            results = []
-            for filename in uploaded_filenames:
-                func = transcriber_live.transcribe_segment.spawn(
-                    filename=filename,
-                )
-                result = func.get()
-                result["filename"] = filename
-                results.append(result)
-
-            return {"results": results} if len(results) > 1 else results[0]
-
-        finally:
-            for filename in uploaded_filenames:
-                try:
-                    file_path = f"{UPLOADS_PATH}/{filename}"
-                    print(f"Deleting file: {file_path}")
-                    os.remove(file_path)
-                except Exception as e:
-                    print(f"Error deleting {filename}: {e}")
-
-            upload_volume.commit()
-
-    @app.post("/v1/audio/transcriptions-from-url", dependencies=[Depends(apikey_auth)])
-    def transcribe_from_url(
-        audio_file_url: str = Body(
-            ..., description="URL of the audio file to transcribe"
-        ),
-        model: str = Body(MODEL_NAME),
-        language: str = Body("en", description="Language code (only 'en' supported)"),
-        timestamp_offset: float = Body(0.0),
-    ):
-        # Parakeet only supports English
-        if language != "en":
-            raise HTTPException(
-                status_code=400,
-                detail=f"Parakeet model only supports English. Got language='{language}'",
-            )
-        unique_filename, audio_suffix = download_audio_to_volume(audio_file_url)
-
-        try:
-            func = transcriber_file.transcribe_segment.spawn(
-                filename=unique_filename,
-                timestamp_offset=timestamp_offset,
-            )
-            result = func.get()
-            return result
-        finally:
-            try:
-                file_path = f"{UPLOADS_PATH}/{unique_filename}"
-                print(f"Deleting file: {file_path}")
-                os.remove(file_path)
-                upload_volume.commit()
-            except Exception as e:
-                print(f"Error cleaning up {unique_filename}: {e}")
-
-    return app
-
-
-class NoStdStreams:
-    def __init__(self):
-        self.devnull = open(os.devnull, "w")
-
-    def __enter__(self):
-        self._stdout, self._stderr = sys.stdout, sys.stderr
-        self._stdout.flush()
-        self._stderr.flush()
-        sys.stdout, sys.stderr = self.devnull, self.devnull
-
-    def __exit__(self, exc_type, exc_value, traceback):
-        sys.stdout, sys.stderr = self._stdout, self._stderr
-        self.devnull.close()

server/migrations/versions/0194f65cd6d3_add_webhook_fields_to_rooms.py

@@ -0,0 +1,36 @@
+"""Add webhook fields to rooms
+
+Revision ID: 0194f65cd6d3
+Revises: 5a8907fd1d78
+Create Date: 2025-08-27 09:03:19.610995
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "0194f65cd6d3"
+down_revision: Union[str, None] = "5a8907fd1d78"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("room", schema=None) as batch_op:
+        batch_op.add_column(sa.Column("webhook_url", sa.String(), nullable=True))
+        batch_op.add_column(sa.Column("webhook_secret", sa.String(), nullable=True))
+
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("room", schema=None) as batch_op:
+        batch_op.drop_column("webhook_secret")
+        batch_op.drop_column("webhook_url")
+
+    # ### end Alembic commands ###

server/migrations/versions/0ce521cda2ee_remove_user_id_from_meeting_table.py

@@ -0,0 +1,36 @@
+"""remove user_id from meeting table
+
+Revision ID: 0ce521cda2ee
+Revises: 6dec9fb5b46c
+Create Date: 2025-09-10 12:40:55.688899
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "0ce521cda2ee"
+down_revision: Union[str, None] = "6dec9fb5b46c"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.drop_column("user_id")
+
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.add_column(
+            sa.Column("user_id", sa.VARCHAR(), autoincrement=False, nullable=True)
+        )
+
+    # ### end Alembic commands ###

server/migrations/versions/1e49625677e4_add_platform_support.py

@@ -0,0 +1,50 @@
+"""add_platform_support
+
+Revision ID: 1e49625677e4
+Revises: dc035ff72fd5
+Create Date: 2025-10-08 13:17:29.943612
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "1e49625677e4"
+down_revision: Union[str, None] = "dc035ff72fd5"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    """Add platform field with default 'whereby' for backward compatibility."""
+    with op.batch_alter_table("room", schema=None) as batch_op:
+        batch_op.add_column(
+            sa.Column(
+                "platform",
+                sa.String(),
+                nullable=False,
+                server_default="whereby",
+            )
+        )
+
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.add_column(
+            sa.Column(
+                "platform",
+                sa.String(),
+                nullable=False,
+                server_default="whereby",
+            )
+        )
+
+
+def downgrade() -> None:
+    """Remove platform field."""
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.drop_column("platform")
+
+    with op.batch_alter_table("room", schema=None) as batch_op:
+        batch_op.drop_column("platform")

server/migrations/versions/2ae3db106d4e_clean_up_orphaned_room_id_references_in_.py

@@ -0,0 +1,32 @@
+"""clean up orphaned room_id references in meeting table
+
+Revision ID: 2ae3db106d4e
+Revises: def1b5867d4c
+Create Date: 2025-09-11 10:35:15.759967
+
+"""
+
+from typing import Sequence, Union
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "2ae3db106d4e"
+down_revision: Union[str, None] = "def1b5867d4c"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # Set room_id to NULL for meetings that reference non-existent rooms
+    op.execute("""
+        UPDATE meeting
+        SET room_id = NULL
+        WHERE room_id IS NOT NULL
+          AND room_id NOT IN (SELECT id FROM room WHERE id IS NOT NULL)
+    """)
+
+
+def downgrade() -> None:
+    # Cannot restore orphaned references - no operation needed
+    pass

server/migrations/versions/5a8907fd1d78_add_cascade_delete_to_meeting_consent_.py

@@ -0,0 +1,50 @@
+"""add cascade delete to meeting consent foreign key
+
+Revision ID: 5a8907fd1d78
+Revises: 0ab2d7ffaa16
+Create Date: 2025-08-26 17:26:50.945491
+
+"""
+
+from typing import Sequence, Union
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "5a8907fd1d78"
+down_revision: Union[str, None] = "0ab2d7ffaa16"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("meeting_consent", schema=None) as batch_op:
+        batch_op.drop_constraint(
+            batch_op.f("meeting_consent_meeting_id_fkey"), type_="foreignkey"
+        )
+        batch_op.create_foreign_key(
+            batch_op.f("meeting_consent_meeting_id_fkey"),
+            "meeting",
+            ["meeting_id"],
+            ["id"],
+            ondelete="CASCADE",
+        )
+
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("meeting_consent", schema=None) as batch_op:
+        batch_op.drop_constraint(
+            batch_op.f("meeting_consent_meeting_id_fkey"), type_="foreignkey"
+        )
+        batch_op.create_foreign_key(
+            batch_op.f("meeting_consent_meeting_id_fkey"),
+            "meeting",
+            ["meeting_id"],
+            ["id"],
+        )
+
+    # ### end Alembic commands ###

server/migrations/versions/6025e9b2bef2_remove_one_active_meeting_per_room_.py

@@ -0,0 +1,53 @@
+"""remove_one_active_meeting_per_room_constraint
+
+Revision ID: 6025e9b2bef2
+Revises: 2ae3db106d4e
+Create Date: 2025-08-18 18:45:44.418392
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "6025e9b2bef2"
+down_revision: Union[str, None] = "2ae3db106d4e"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # Remove the unique constraint that prevents multiple active meetings per room
+    # This is needed to support calendar integration with overlapping meetings
+    # Check if index exists before trying to drop it
+    from alembic import context
+
+    if context.get_context().dialect.name == "postgresql":
+        conn = op.get_bind()
+        result = conn.execute(
+            sa.text(
+                "SELECT 1 FROM pg_indexes WHERE indexname = 'idx_one_active_meeting_per_room'"
+            )
+        )
+        if result.fetchone():
+            op.drop_index("idx_one_active_meeting_per_room", table_name="meeting")
+    else:
+        # For SQLite, just try to drop it
+        try:
+            op.drop_index("idx_one_active_meeting_per_room", table_name="meeting")
+        except:
+            pass
+
+
+def downgrade() -> None:
+    # Restore the unique constraint
+    op.create_index(
+        "idx_one_active_meeting_per_room",
+        "meeting",
+        ["room_id"],
+        unique=True,
+        postgresql_where=sa.text("is_active = true"),
+        sqlite_where=sa.text("is_active = 1"),
+    )

server/migrations/versions/61882a919591_webhook_url_and_secret_null_by_default.py

@@ -0,0 +1,28 @@
+"""webhook url and secret null by default
+
+
+Revision ID: 61882a919591
+Revises: 0194f65cd6d3
+Create Date: 2025-08-29 11:46:36.738091
+
+"""
+
+from typing import Sequence, Union
+
+# revision identifiers, used by Alembic.
+revision: str = "61882a919591"
+down_revision: Union[str, None] = "0194f65cd6d3"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    pass
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    pass
+    # ### end Alembic commands ###

server/migrations/versions/6dec9fb5b46c_make_meeting_room_id_required_and_add_.py

@@ -0,0 +1,35 @@
+"""make meeting room_id required and add foreign key
+
+Revision ID: 6dec9fb5b46c
+Revises: 61882a919591
+Create Date: 2025-09-10 10:47:06.006819
+
+"""
+
+from typing import Sequence, Union
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "6dec9fb5b46c"
+down_revision: Union[str, None] = "61882a919591"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.create_foreign_key(
+            None, "room", ["room_id"], ["id"], ondelete="CASCADE"
+        )
+
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.drop_constraint("meeting_room_id_fkey", type_="foreignkey")
+
+    # ### end Alembic commands ###

server/migrations/versions/d4a1c446458c_add_grace_period_fields_to_meeting.py

@@ -0,0 +1,34 @@
+"""add_grace_period_fields_to_meeting
+
+Revision ID: d4a1c446458c
+Revises: 6025e9b2bef2
+Create Date: 2025-08-18 18:50:37.768052
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "d4a1c446458c"
+down_revision: Union[str, None] = "6025e9b2bef2"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # Add fields to track when participants left for grace period logic
+    op.add_column(
+        "meeting", sa.Column("last_participant_left_at", sa.DateTime(timezone=True))
+    )
+    op.add_column(
+        "meeting",
+        sa.Column("grace_period_minutes", sa.Integer, server_default=sa.text("15")),
+    )
+
+
+def downgrade() -> None:
+    op.drop_column("meeting", "grace_period_minutes")
+    op.drop_column("meeting", "last_participant_left_at")

server/migrations/versions/d8e204bbf615_add_calendar.py

@@ -0,0 +1,129 @@
+"""add calendar
+
+Revision ID: d8e204bbf615
+Revises: d4a1c446458c
+Create Date: 2025-09-10 19:56:22.295756
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+from sqlalchemy.dialects import postgresql
+
+# revision identifiers, used by Alembic.
+revision: str = "d8e204bbf615"
+down_revision: Union[str, None] = "d4a1c446458c"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    op.create_table(
+        "calendar_event",
+        sa.Column("id", sa.String(), nullable=False),
+        sa.Column("room_id", sa.String(), nullable=False),
+        sa.Column("ics_uid", sa.Text(), nullable=False),
+        sa.Column("title", sa.Text(), nullable=True),
+        sa.Column("description", sa.Text(), nullable=True),
+        sa.Column("start_time", sa.DateTime(timezone=True), nullable=False),
+        sa.Column("end_time", sa.DateTime(timezone=True), nullable=False),
+        sa.Column("attendees", postgresql.JSONB(astext_type=sa.Text()), nullable=True),
+        sa.Column("location", sa.Text(), nullable=True),
+        sa.Column("ics_raw_data", sa.Text(), nullable=True),
+        sa.Column("last_synced", sa.DateTime(timezone=True), nullable=False),
+        sa.Column(
+            "is_deleted", sa.Boolean(), server_default=sa.text("false"), nullable=False
+        ),
+        sa.Column("created_at", sa.DateTime(timezone=True), nullable=False),
+        sa.Column("updated_at", sa.DateTime(timezone=True), nullable=False),
+        sa.ForeignKeyConstraint(
+            ["room_id"],
+            ["room.id"],
+            name="fk_calendar_event_room_id",
+            ondelete="CASCADE",
+        ),
+        sa.PrimaryKeyConstraint("id"),
+        sa.UniqueConstraint("room_id", "ics_uid", name="uq_room_calendar_event"),
+    )
+    with op.batch_alter_table("calendar_event", schema=None) as batch_op:
+        batch_op.create_index(
+            "idx_calendar_event_deleted",
+            ["is_deleted"],
+            unique=False,
+            postgresql_where=sa.text("NOT is_deleted"),
+        )
+        batch_op.create_index(
+            "idx_calendar_event_room_start", ["room_id", "start_time"], unique=False
+        )
+
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.add_column(sa.Column("calendar_event_id", sa.String(), nullable=True))
+        batch_op.add_column(
+            sa.Column(
+                "calendar_metadata",
+                postgresql.JSONB(astext_type=sa.Text()),
+                nullable=True,
+            )
+        )
+        batch_op.create_index(
+            "idx_meeting_calendar_event", ["calendar_event_id"], unique=False
+        )
+        batch_op.create_foreign_key(
+            "fk_meeting_calendar_event_id",
+            "calendar_event",
+            ["calendar_event_id"],
+            ["id"],
+            ondelete="SET NULL",
+        )
+
+    with op.batch_alter_table("room", schema=None) as batch_op:
+        batch_op.add_column(sa.Column("ics_url", sa.Text(), nullable=True))
+        batch_op.add_column(
+            sa.Column(
+                "ics_fetch_interval", sa.Integer(), server_default="300", nullable=True
+            )
+        )
+        batch_op.add_column(
+            sa.Column(
+                "ics_enabled",
+                sa.Boolean(),
+                server_default=sa.text("false"),
+                nullable=False,
+            )
+        )
+        batch_op.add_column(
+            sa.Column("ics_last_sync", sa.DateTime(timezone=True), nullable=True)
+        )
+        batch_op.add_column(sa.Column("ics_last_etag", sa.Text(), nullable=True))
+        batch_op.create_index("idx_room_ics_enabled", ["ics_enabled"], unique=False)
+
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("room", schema=None) as batch_op:
+        batch_op.drop_index("idx_room_ics_enabled")
+        batch_op.drop_column("ics_last_etag")
+        batch_op.drop_column("ics_last_sync")
+        batch_op.drop_column("ics_enabled")
+        batch_op.drop_column("ics_fetch_interval")
+        batch_op.drop_column("ics_url")
+
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.drop_constraint("fk_meeting_calendar_event_id", type_="foreignkey")
+        batch_op.drop_index("idx_meeting_calendar_event")
+        batch_op.drop_column("calendar_metadata")
+        batch_op.drop_column("calendar_event_id")
+
+    with op.batch_alter_table("calendar_event", schema=None) as batch_op:
+        batch_op.drop_index("idx_calendar_event_room_start")
+        batch_op.drop_index(
+            "idx_calendar_event_deleted", postgresql_where=sa.text("NOT is_deleted")
+        )
+
+    op.drop_table("calendar_event")
+    # ### end Alembic commands ###

server/migrations/versions/dc035ff72fd5_remove_grace_period_fields.py

@@ -0,0 +1,43 @@
+"""remove_grace_period_fields
+
+Revision ID: dc035ff72fd5
+Revises: d8e204bbf615
+Create Date: 2025-09-11 10:36:45.197588
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "dc035ff72fd5"
+down_revision: Union[str, None] = "d8e204bbf615"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # Remove grace period columns from meeting table
+    op.drop_column("meeting", "last_participant_left_at")
+    op.drop_column("meeting", "grace_period_minutes")
+
+
+def downgrade() -> None:
+    # Add back grace period columns to meeting table
+    op.add_column(
+        "meeting",
+        sa.Column(
+            "last_participant_left_at", sa.DateTime(timezone=True), nullable=True
+        ),
+    )
+    op.add_column(
+        "meeting",
+        sa.Column(
+            "grace_period_minutes",
+            sa.Integer(),
+            server_default=sa.text("15"),
+            nullable=True,
+        ),
+    )

server/migrations/versions/def1b5867d4c_make_meeting_room_id_nullable_but_keep_.py

@@ -0,0 +1,34 @@
+"""make meeting room_id nullable but keep foreign key
+
+Revision ID: def1b5867d4c
+Revises: 0ce521cda2ee
+Create Date: 2025-09-11 09:42:18.697264
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "def1b5867d4c"
+down_revision: Union[str, None] = "0ce521cda2ee"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.alter_column("room_id", existing_type=sa.VARCHAR(), nullable=True)
+
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.alter_column("room_id", existing_type=sa.VARCHAR(), nullable=False)
+
+    # ### end Alembic commands ###

server/pyproject.toml

@@ -12,7 +12,6 @@ dependencies = [
     "requests>=2.31.0",
     "aiortc>=1.5.0",
     "sortedcontainers>=2.4.0",
-    "loguru>=0.7.0",
     "pydantic-settings>=2.0.2",
     "structlog>=23.1.0",
     "uvicorn[standard]>=0.23.1",
@@ -27,7 +26,6 @@ dependencies = [
     "prometheus-fastapi-instrumentator>=6.1.0",
     "sentencepiece>=0.1.99",
     "protobuf>=4.24.3",
-    "profanityfilter>=2.0.6",
     "celery>=5.3.4",
     "redis>=5.0.1",
     "python-jose[cryptography]>=3.3.0",
@@ -40,6 +38,7 @@ dependencies = [
     "llama-index-llms-openai-like>=0.4.0",
     "pytest-env>=1.1.5",
     "webvtt-py>=0.5.0",
+    "icalendar>=6.0.0",
 ]
 
 [dependency-groups]
@@ -113,13 +112,14 @@ source = ["reflector"]
 [tool.pytest_env]
 ENVIRONMENT = "pytest"
 DATABASE_URL = "postgresql://test_user:test_password@localhost:15432/reflector_test"
+AUTH_BACKEND = "jwt"
 
 [tool.pytest.ini_options]
 addopts = "-ra -q --disable-pytest-warnings --cov --cov-report html -v"
 testpaths = ["tests"]
 asyncio_mode = "auto"
 markers = [
-    "gpu_modal: mark test to run only with GPU Modal endpoints (deselect with '-m \"not gpu_modal\"')",
+    "model_api: tests for the unified model-serving HTTP API (backend- and hardware-agnostic)",
 ]
 
 [tool.ruff.lint]
@@ -131,7 +131,7 @@ select = [
 
 [tool.ruff.lint.per-file-ignores]
 "reflector/processors/summary/summary_builder.py" = ["E501"]
-"gpu/**.py" = ["PLC0415"]
+"gpu/modal_deployments/**.py" = ["PLC0415"]
 "reflector/tools/**.py" = ["PLC0415"]
 "migrations/versions/**.py" = ["PLC0415"]
 "tests/**.py" = ["PLC0415"]

server/reflector/app.py

@@ -12,6 +12,7 @@ from reflector.events import subscribers_shutdown, subscribers_startup
 from reflector.logger import logger
 from reflector.metrics import metrics_init
 from reflector.settings import settings
+from reflector.views.daily import router as daily_router
 from reflector.views.meetings import router as meetings_router
 from reflector.views.rooms import router as rooms_router
 from reflector.views.rtc_offer import router as rtc_offer_router
@@ -26,6 +27,7 @@ from reflector.views.transcripts_upload import router as transcripts_upload_rout
 from reflector.views.transcripts_webrtc import router as transcripts_webrtc_router
 from reflector.views.transcripts_websocket import router as transcripts_websocket_router
 from reflector.views.user import router as user_router
+from reflector.views.user_websocket import router as user_ws_router
 from reflector.views.whereby import router as whereby_router
 from reflector.views.zulip import router as zulip_router
 
@@ -65,6 +67,12 @@ app.add_middleware(
     allow_headers=["*"],
 )
 
+
+@app.get("/health")
+async def health():
+    return {"status": "healthy"}
+
+
 # metrics
 instrumentator = Instrumentator(
     excluded_handlers=["/docs", "/metrics"],
@@ -84,8 +92,10 @@ app.include_router(transcripts_websocket_router, prefix="/v1")
 app.include_router(transcripts_webrtc_router, prefix="/v1")
 app.include_router(transcripts_process_router, prefix="/v1")
 app.include_router(user_router, prefix="/v1")
+app.include_router(user_ws_router, prefix="/v1")
 app.include_router(zulip_router, prefix="/v1")
 app.include_router(whereby_router, prefix="/v1")
+app.include_router(daily_router, prefix="/v1/daily")
 add_pagination(app)
 
 # prepare celery

server/reflector/asynctask.py

@@ -0,0 +1,27 @@
+import asyncio
+import functools
+
+from reflector.db import get_database
+
+
+def asynctask(f):
+    @functools.wraps(f)
+    def wrapper(*args, **kwargs):
+        async def run_with_db():
+            database = get_database()
+            await database.connect()
+            try:
+                return await f(*args, **kwargs)
+            finally:
+                await database.disconnect()
+
+        coro = run_with_db()
+        try:
+            loop = asyncio.get_running_loop()
+        except RuntimeError:
+            loop = None
+        if loop and loop.is_running():
+            return loop.run_until_complete(coro)
+        return asyncio.run(coro)
+
+    return wrapper

server/reflector/auth/auth_jwt.py

@@ -67,7 +67,8 @@ def current_user(
     try:
         payload = jwtauth.verify_token(token)
         sub = payload["sub"]
-        return UserInfo(sub=sub)
+        email = payload["email"]
+        return UserInfo(sub=sub, email=email)
     except JWTError as e:
         logger.error(f"JWT error: {e}")
         raise HTTPException(status_code=401, detail="Invalid authentication")

server/reflector/db/__init__.py

@@ -24,6 +24,7 @@ def get_database() -> databases.Database:
 
 
 # import models
+import reflector.db.calendar_events  # noqa
 import reflector.db.meetings  # noqa
 import reflector.db.recordings  # noqa
 import reflector.db.rooms  # noqa

server/reflector/db/calendar_events.py

@@ -0,0 +1,187 @@
+from datetime import datetime, timedelta, timezone
+from typing import Any
+
+import sqlalchemy as sa
+from pydantic import BaseModel, Field
+from sqlalchemy.dialects.postgresql import JSONB
+
+from reflector.db import get_database, metadata
+from reflector.utils import generate_uuid4
+
+calendar_events = sa.Table(
+    "calendar_event",
+    metadata,
+    sa.Column("id", sa.String, primary_key=True),
+    sa.Column(
+        "room_id",
+        sa.String,
+        sa.ForeignKey("room.id", ondelete="CASCADE", name="fk_calendar_event_room_id"),
+        nullable=False,
+    ),
+    sa.Column("ics_uid", sa.Text, nullable=False),
+    sa.Column("title", sa.Text),
+    sa.Column("description", sa.Text),
+    sa.Column("start_time", sa.DateTime(timezone=True), nullable=False),
+    sa.Column("end_time", sa.DateTime(timezone=True), nullable=False),
+    sa.Column("attendees", JSONB),
+    sa.Column("location", sa.Text),
+    sa.Column("ics_raw_data", sa.Text),
+    sa.Column("last_synced", sa.DateTime(timezone=True), nullable=False),
+    sa.Column("is_deleted", sa.Boolean, nullable=False, server_default=sa.false()),
+    sa.Column("created_at", sa.DateTime(timezone=True), nullable=False),
+    sa.Column("updated_at", sa.DateTime(timezone=True), nullable=False),
+    sa.UniqueConstraint("room_id", "ics_uid", name="uq_room_calendar_event"),
+    sa.Index("idx_calendar_event_room_start", "room_id", "start_time"),
+    sa.Index(
+        "idx_calendar_event_deleted",
+        "is_deleted",
+        postgresql_where=sa.text("NOT is_deleted"),
+    ),
+)
+
+
+class CalendarEvent(BaseModel):
+    id: str = Field(default_factory=generate_uuid4)
+    room_id: str
+    ics_uid: str
+    title: str | None = None
+    description: str | None = None
+    start_time: datetime
+    end_time: datetime
+    attendees: list[dict[str, Any]] | None = None
+    location: str | None = None
+    ics_raw_data: str | None = None
+    last_synced: datetime = Field(default_factory=lambda: datetime.now(timezone.utc))
+    is_deleted: bool = False
+    created_at: datetime = Field(default_factory=lambda: datetime.now(timezone.utc))
+    updated_at: datetime = Field(default_factory=lambda: datetime.now(timezone.utc))
+
+
+class CalendarEventController:
+    async def get_by_room(
+        self,
+        room_id: str,
+        include_deleted: bool = False,
+        start_after: datetime | None = None,
+        end_before: datetime | None = None,
+    ) -> list[CalendarEvent]:
+        query = calendar_events.select().where(calendar_events.c.room_id == room_id)
+
+        if not include_deleted:
+            query = query.where(calendar_events.c.is_deleted == False)
+
+        if start_after:
+            query = query.where(calendar_events.c.start_time >= start_after)
+
+        if end_before:
+            query = query.where(calendar_events.c.end_time <= end_before)
+
+        query = query.order_by(calendar_events.c.start_time.asc())
+
+        results = await get_database().fetch_all(query)
+        return [CalendarEvent(**result) for result in results]
+
+    async def get_upcoming(
+        self, room_id: str, minutes_ahead: int = 120
+    ) -> list[CalendarEvent]:
+        """Get upcoming events for a room within the specified minutes, including currently happening events."""
+        now = datetime.now(timezone.utc)
+        future_time = now + timedelta(minutes=minutes_ahead)
+
+        query = (
+            calendar_events.select()
+            .where(
+                sa.and_(
+                    calendar_events.c.room_id == room_id,
+                    calendar_events.c.is_deleted == False,
+                    calendar_events.c.start_time <= future_time,
+                    calendar_events.c.end_time >= now,
+                )
+            )
+            .order_by(calendar_events.c.start_time.asc())
+        )
+
+        results = await get_database().fetch_all(query)
+        return [CalendarEvent(**result) for result in results]
+
+    async def get_by_id(self, event_id: str) -> CalendarEvent | None:
+        query = calendar_events.select().where(calendar_events.c.id == event_id)
+        result = await get_database().fetch_one(query)
+        return CalendarEvent(**result) if result else None
+
+    async def get_by_ics_uid(self, room_id: str, ics_uid: str) -> CalendarEvent | None:
+        query = calendar_events.select().where(
+            sa.and_(
+                calendar_events.c.room_id == room_id,
+                calendar_events.c.ics_uid == ics_uid,
+            )
+        )
+        result = await get_database().fetch_one(query)
+        return CalendarEvent(**result) if result else None
+
+    async def upsert(self, event: CalendarEvent) -> CalendarEvent:
+        existing = await self.get_by_ics_uid(event.room_id, event.ics_uid)
+
+        if existing:
+            event.id = existing.id
+            event.created_at = existing.created_at
+            event.updated_at = datetime.now(timezone.utc)
+
+            query = (
+                calendar_events.update()
+                .where(calendar_events.c.id == existing.id)
+                .values(**event.model_dump())
+            )
+        else:
+            query = calendar_events.insert().values(**event.model_dump())
+
+        await get_database().execute(query)
+        return event
+
+    async def soft_delete_missing(
+        self, room_id: str, current_ics_uids: list[str]
+    ) -> int:
+        """Soft delete future events that are no longer in the calendar."""
+        now = datetime.now(timezone.utc)
+
+        select_query = calendar_events.select().where(
+            sa.and_(
+                calendar_events.c.room_id == room_id,
+                calendar_events.c.start_time > now,
+                calendar_events.c.is_deleted == False,
+                calendar_events.c.ics_uid.notin_(current_ics_uids)
+                if current_ics_uids
+                else True,
+            )
+        )
+
+        to_delete = await get_database().fetch_all(select_query)
+        delete_count = len(to_delete)
+
+        if delete_count > 0:
+            update_query = (
+                calendar_events.update()
+                .where(
+                    sa.and_(
+                        calendar_events.c.room_id == room_id,
+                        calendar_events.c.start_time > now,
+                        calendar_events.c.is_deleted == False,
+                        calendar_events.c.ics_uid.notin_(current_ics_uids)
+                        if current_ics_uids
+                        else True,
+                    )
+                )
+                .values(is_deleted=True, updated_at=now)
+            )
+
+            await get_database().execute(update_query)
+
+        return delete_count
+
+    async def delete_by_room(self, room_id: str) -> int:
+        query = calendar_events.delete().where(calendar_events.c.room_id == room_id)
+        result = await get_database().execute(query)
+        return result.rowcount
+
+
+calendar_events_controller = CalendarEventController()

server/reflector/db/meetings.py

@@ -1,12 +1,13 @@
 from datetime import datetime
-from typing import Literal
+from typing import Any, Literal
 
 import sqlalchemy as sa
-from fastapi import HTTPException
 from pydantic import BaseModel, Field
+from sqlalchemy.dialects.postgresql import JSONB
 
 from reflector.db import get_database, metadata
 from reflector.db.rooms import Room
+from reflector.platform_types import Platform
 from reflector.utils import generate_uuid4
 
 meetings = sa.Table(
@@ -18,8 +19,12 @@ meetings = sa.Table(
     sa.Column("host_room_url", sa.String),
     sa.Column("start_date", sa.DateTime(timezone=True)),
     sa.Column("end_date", sa.DateTime(timezone=True)),
-    sa.Column("user_id", sa.String),
-    sa.Column("room_id", sa.String),
+    sa.Column(
+        "room_id",
+        sa.String,
+        sa.ForeignKey("room.id", ondelete="CASCADE"),
+        nullable=True,
+    ),
     sa.Column("is_locked", sa.Boolean, nullable=False, server_default=sa.false()),
     sa.Column("room_mode", sa.String, nullable=False, server_default="normal"),
     sa.Column("recording_type", sa.String, nullable=False, server_default="cloud"),
@@ -41,20 +46,36 @@ meetings = sa.Table(
         nullable=False,
         server_default=sa.true(),
     ),
-    sa.Index("idx_meeting_room_id", "room_id"),
-    sa.Index(
-        "idx_one_active_meeting_per_room",
-        "room_id",
-        unique=True,
-        postgresql_where=sa.text("is_active = true"),
+    sa.Column(
+        "calendar_event_id",
+        sa.String,
+        sa.ForeignKey(
+            "calendar_event.id",
+            ondelete="SET NULL",
+            name="fk_meeting_calendar_event_id",
+        ),
     ),
+    sa.Column("calendar_metadata", JSONB),
+    sa.Column(
+        "platform",
+        sa.String,
+        nullable=False,
+        server_default="whereby",
+    ),
+    sa.Index("idx_meeting_room_id", "room_id"),
+    sa.Index("idx_meeting_calendar_event", "calendar_event_id"),
 )
 
 meeting_consent = sa.Table(
     "meeting_consent",
     metadata,
     sa.Column("id", sa.String, primary_key=True),
-    sa.Column("meeting_id", sa.String, sa.ForeignKey("meeting.id"), nullable=False),
+    sa.Column(
+        "meeting_id",
+        sa.String,
+        sa.ForeignKey("meeting.id", ondelete="CASCADE"),
+        nullable=False,
+    ),
     sa.Column("user_id", sa.String),
     sa.Column("consent_given", sa.Boolean, nullable=False),
     sa.Column("consent_timestamp", sa.DateTime(timezone=True), nullable=False),
@@ -76,15 +97,18 @@ class Meeting(BaseModel):
     host_room_url: str
     start_date: datetime
     end_date: datetime
-    user_id: str | None = None
-    room_id: str | None = None
+    room_id: str | None
     is_locked: bool = False
     room_mode: Literal["normal", "group"] = "normal"
     recording_type: Literal["none", "local", "cloud"] = "cloud"
-    recording_trigger: Literal[
+    recording_trigger: Literal[  # whereby-specific
         "none", "prompt", "automatic", "automatic-2nd-participant"
     ] = "automatic-2nd-participant"
     num_clients: int = 0
+    is_active: bool = True
+    calendar_event_id: str | None = None
+    calendar_metadata: dict[str, Any] | None = None
+    platform: Platform = "whereby"
 
 
 class MeetingController:
@@ -96,12 +120,11 @@ class MeetingController:
         host_room_url: str,
         start_date: datetime,
         end_date: datetime,
-        user_id: str,
         room: Room,
+        calendar_event_id: str | None = None,
+        calendar_metadata: dict[str, Any] | None = None,
+        platform: Platform = "whereby",
     ):
-        """
-        Create a new meeting
-        """
         meeting = Meeting(
             id=id,
             room_name=room_name,
@@ -109,41 +132,47 @@ class MeetingController:
             host_room_url=host_room_url,
             start_date=start_date,
             end_date=end_date,
-            user_id=user_id,
             room_id=room.id,
             is_locked=room.is_locked,
             room_mode=room.room_mode,
             recording_type=room.recording_type,
             recording_trigger=room.recording_trigger,
+            calendar_event_id=calendar_event_id,
+            calendar_metadata=calendar_metadata,
+            platform=platform,
         )
         query = meetings.insert().values(**meeting.model_dump())
         await get_database().execute(query)
         return meeting
 
     async def get_all_active(self) -> list[Meeting]:
-        """
-        Get active meetings.
-        """
         query = meetings.select().where(meetings.c.is_active)
         return await get_database().fetch_all(query)
 
     async def get_by_room_name(
         self,
         room_name: str,
-    ) -> Meeting:
+    ) -> Meeting | None:
         """
         Get a meeting by room name.
+        For backward compatibility, returns the most recent meeting.
         """
-        query = meetings.select().where(meetings.c.room_name == room_name)
+        end_date = getattr(meetings.c, "end_date")
+        query = (
+            meetings.select()
+            .where(meetings.c.room_name == room_name)
+            .order_by(end_date.desc())
+        )
         result = await get_database().fetch_one(query)
         if not result:
             return None
 
         return Meeting(**result)
 
-    async def get_active(self, room: Room, current_time: datetime) -> Meeting:
+    async def get_active(self, room: Room, current_time: datetime) -> Meeting | None:
         """
         Get latest active meeting for a room.
+        For backward compatibility, returns the most recent active meeting.
         """
         end_date = getattr(meetings.c, "end_date")
         query = (
@@ -163,37 +192,85 @@ class MeetingController:
 
         return Meeting(**result)
 
+    async def get_all_active_for_room(
+        self, room: Room, current_time: datetime
+    ) -> list[Meeting]:
+        end_date = getattr(meetings.c, "end_date")
+        query = (
+            meetings.select()
+            .where(
+                sa.and_(
+                    meetings.c.room_id == room.id,
+                    meetings.c.end_date > current_time,
+                    meetings.c.is_active,
+                )
+            )
+            .order_by(end_date.desc())
+        )
+        results = await get_database().fetch_all(query)
+        return [Meeting(**result) for result in results]
+
+    async def get_active_by_calendar_event(
+        self, room: Room, calendar_event_id: str, current_time: datetime
+    ) -> Meeting | None:
+        """
+        Get active meeting for a specific calendar event.
+        """
+        query = meetings.select().where(
+            sa.and_(
+                meetings.c.room_id == room.id,
+                meetings.c.calendar_event_id == calendar_event_id,
+                meetings.c.end_date > current_time,
+                meetings.c.is_active,
+            )
+        )
+        result = await get_database().fetch_one(query)
+        if not result:
+            return None
+        return Meeting(**result)
+
     async def get_by_id(self, meeting_id: str, **kwargs) -> Meeting | None:
-        """
-        Get a meeting by id
-        """
         query = meetings.select().where(meetings.c.id == meeting_id)
         result = await get_database().fetch_one(query)
         if not result:
             return None
         return Meeting(**result)
 
-    async def get_by_id_for_http(self, meeting_id: str, user_id: str | None) -> Meeting:
-        """
-        Get a meeting by ID for HTTP request.
-
-        If not found, it will raise a 404 error.
-        """
-        query = meetings.select().where(meetings.c.id == meeting_id)
+    async def get_by_calendar_event(self, calendar_event_id: str) -> Meeting | None:
+        query = meetings.select().where(
+            meetings.c.calendar_event_id == calendar_event_id
+        )
         result = await get_database().fetch_one(query)
         if not result:
-            raise HTTPException(status_code=404, detail="Meeting not found")
-
-        meeting = Meeting(**result)
-        if result["user_id"] != user_id:
-            meeting.host_room_url = ""
-
-        return meeting
+            return None
+        return Meeting(**result)
 
     async def update_meeting(self, meeting_id: str, **kwargs):
         query = meetings.update().where(meetings.c.id == meeting_id).values(**kwargs)
         await get_database().execute(query)
 
+    async def increment_num_clients(self, meeting_id: str):
+        """Atomically increment participant count."""
+        query = (
+            meetings.update()
+            .where(meetings.c.id == meeting_id)
+            .values(num_clients=meetings.c.num_clients + 1)
+        )
+        await get_database().execute(query)
+
+    async def decrement_num_clients(self, meeting_id: str):
+        """Atomically decrement participant count (min 0)."""
+        query = (
+            meetings.update()
+            .where(meetings.c.id == meeting_id)
+            .values(
+                num_clients=sa.case(
+                    (meetings.c.num_clients > 0, meetings.c.num_clients - 1), else_=0
+                )
+            )
+        )
+        await get_database().execute(query)
+
 
 class MeetingConsentController:
     async def get_by_meeting_id(self, meeting_id: str) -> list[MeetingConsent]:
@@ -214,10 +291,9 @@ class MeetingConsentController:
         result = await get_database().fetch_one(query)
         if result is None:
             return None
-        return MeetingConsent(**result) if result else None
+        return MeetingConsent(**result)
 
     async def upsert(self, consent: MeetingConsent) -> MeetingConsent:
-        """Create new consent or update existing one for authenticated users"""
         if consent.user_id:
             # For authenticated users, check if consent already exists
             # not transactional but we're ok with that; the consents ain't deleted anyways

server/reflector/db/rooms.py

@@ -1,6 +1,7 @@
+import secrets
 from datetime import datetime, timezone
 from sqlite3 import IntegrityError
-from typing import Literal
+from typing import Literal, Optional
 
 import sqlalchemy
 from fastapi import HTTPException
@@ -8,6 +9,7 @@ from pydantic import BaseModel, Field
 from sqlalchemy.sql import false, or_
 
 from reflector.db import get_database, metadata
+from reflector.platform_types import Platform
 from reflector.utils import generate_uuid4
 
 rooms = sqlalchemy.Table(
@@ -40,7 +42,23 @@ rooms = sqlalchemy.Table(
     sqlalchemy.Column(
         "is_shared", sqlalchemy.Boolean, nullable=False, server_default=false()
     ),
+    sqlalchemy.Column("webhook_url", sqlalchemy.String, nullable=True),
+    sqlalchemy.Column("webhook_secret", sqlalchemy.String, nullable=True),
+    sqlalchemy.Column("ics_url", sqlalchemy.Text),
+    sqlalchemy.Column("ics_fetch_interval", sqlalchemy.Integer, server_default="300"),
+    sqlalchemy.Column(
+        "ics_enabled", sqlalchemy.Boolean, nullable=False, server_default=false()
+    ),
+    sqlalchemy.Column("ics_last_sync", sqlalchemy.DateTime(timezone=True)),
+    sqlalchemy.Column("ics_last_etag", sqlalchemy.Text),
+    sqlalchemy.Column(
+        "platform",
+        sqlalchemy.String,
+        nullable=False,
+        server_default="whereby",
+    ),
     sqlalchemy.Index("idx_room_is_shared", "is_shared"),
+    sqlalchemy.Index("idx_room_ics_enabled", "ics_enabled"),
 )
 
 
@@ -55,10 +73,18 @@ class Room(BaseModel):
     is_locked: bool = False
     room_mode: Literal["normal", "group"] = "normal"
     recording_type: Literal["none", "local", "cloud"] = "cloud"
-    recording_trigger: Literal[
+    recording_trigger: Literal[  # whereby-specific
         "none", "prompt", "automatic", "automatic-2nd-participant"
     ] = "automatic-2nd-participant"
     is_shared: bool = False
+    webhook_url: str | None = None
+    webhook_secret: str | None = None
+    ics_url: str | None = None
+    ics_fetch_interval: int = 300
+    ics_enabled: bool = False
+    ics_last_sync: datetime | None = None
+    ics_last_etag: str | None = None
+    platform: Platform = "whereby"
 
 
 class RoomController:
@@ -107,10 +133,19 @@ class RoomController:
         recording_type: str,
         recording_trigger: str,
         is_shared: bool,
+        webhook_url: str = "",
+        webhook_secret: str = "",
+        ics_url: str | None = None,
+        ics_fetch_interval: int = 300,
+        ics_enabled: bool = False,
+        platform: Optional[Platform] = None,
     ):
         """
         Add a new room
         """
+        if webhook_url and not webhook_secret:
+            webhook_secret = secrets.token_urlsafe(32)
+
         room = Room(
             name=name,
             user_id=user_id,
@@ -122,6 +157,12 @@ class RoomController:
             recording_type=recording_type,
             recording_trigger=recording_trigger,
             is_shared=is_shared,
+            webhook_url=webhook_url,
+            webhook_secret=webhook_secret,
+            ics_url=ics_url,
+            ics_fetch_interval=ics_fetch_interval,
+            ics_enabled=ics_enabled,
+            platform=platform or "whereby",
         )
         query = rooms.insert().values(**room.model_dump())
         try:
@@ -134,6 +175,9 @@ class RoomController:
         """
         Update a room fields with key/values in values
         """
+        if values.get("webhook_url") and not values.get("webhook_secret"):
+            values["webhook_secret"] = secrets.token_urlsafe(32)
+
         query = rooms.update().where(rooms.c.id == room.id).values(**values)
         try:
             await get_database().execute(query)
@@ -183,6 +227,13 @@ class RoomController:
 
         return room
 
+    async def get_ics_enabled(self) -> list[Room]:
+        query = rooms.select().where(
+            rooms.c.ics_enabled == True, rooms.c.ics_url != None
+        )
+        results = await get_database().fetch_all(query)
+        return [Room(**result) for result in results]
+
     async def remove_by_id(
         self,
         room_id: str,

server/reflector/db/search.py

@@ -8,12 +8,14 @@ from typing import Annotated, Any, Dict, Iterator
 
 import sqlalchemy
 import webvtt
+from databases.interfaces import Record as DbRecord
 from fastapi import HTTPException
 from pydantic import (
     BaseModel,
     Field,
     NonNegativeFloat,
     NonNegativeInt,
+    TypeAdapter,
     ValidationError,
     constr,
     field_serializer,
@@ -21,9 +23,10 @@ from pydantic import (
 
 from reflector.db import get_database
 from reflector.db.rooms import rooms
-from reflector.db.transcripts import SourceKind, transcripts
+from reflector.db.transcripts import SourceKind, TranscriptStatus, transcripts
 from reflector.db.utils import is_postgresql
 from reflector.logger import logger
+from reflector.utils.string import NonEmptyString, try_parse_non_empty_string
 
 DEFAULT_SEARCH_LIMIT = 20
 SNIPPET_CONTEXT_LENGTH = 50  # Characters before/after match to include
@@ -31,12 +34,13 @@ DEFAULT_SNIPPET_MAX_LENGTH = NonNegativeInt(150)
 DEFAULT_MAX_SNIPPETS = NonNegativeInt(3)
 LONG_SUMMARY_MAX_SNIPPETS = 2
 
-SearchQueryBase = constr(min_length=0, strip_whitespace=True)
+SearchQueryBase = constr(min_length=1, strip_whitespace=True)
 SearchLimitBase = Annotated[int, Field(ge=1, le=100)]
 SearchOffsetBase = Annotated[int, Field(ge=0)]
 SearchTotalBase = Annotated[int, Field(ge=0)]
 
 SearchQuery = Annotated[SearchQueryBase, Field(description="Search query text")]
+search_query_adapter = TypeAdapter(SearchQuery)
 SearchLimit = Annotated[SearchLimitBase, Field(description="Results per page")]
 SearchOffset = Annotated[
     SearchOffsetBase, Field(description="Number of results to skip")
@@ -88,7 +92,7 @@ class WebVTTProcessor:
     @staticmethod
     def generate_snippets(
         webvtt_content: WebVTTContent,
-        query: str,
+        query: SearchQuery,
         max_snippets: NonNegativeInt = DEFAULT_MAX_SNIPPETS,
     ) -> list[str]:
         """Generate snippets from WebVTT content."""
@@ -125,7 +129,7 @@ class SnippetCandidate:
 class SearchParameters(BaseModel):
     """Validated search parameters for full-text search."""
 
-    query_text: SearchQuery
+    query_text: SearchQuery | None = None
     limit: SearchLimit = DEFAULT_SEARCH_LIMIT
     offset: SearchOffset = 0
     user_id: str | None = None
@@ -157,7 +161,7 @@ class SearchResult(BaseModel):
     room_name: str | None = None
     source_kind: SourceKind
     created_at: datetime
-    status: str = Field(..., min_length=1)
+    status: TranscriptStatus = Field(..., min_length=1)
     rank: float = Field(..., ge=0, le=1)
     duration: NonNegativeFloat | None = Field(..., description="Duration in seconds")
     search_snippets: list[str] = Field(
@@ -199,15 +203,13 @@ class SnippetGenerator:
             prev_start = start
 
     @staticmethod
-    def count_matches(text: str, query: str) -> NonNegativeInt:
+    def count_matches(text: str, query: SearchQuery) -> NonNegativeInt:
         """Count total number of matches for a query in text."""
         ZERO = NonNegativeInt(0)
         if not text:
             logger.warning("Empty text for search query in count_matches")
             return ZERO
-        if not query:
-            logger.warning("Empty query for search text in count_matches")
-            return ZERO
+        assert query is not None
         return NonNegativeInt(
             sum(1 for _ in SnippetGenerator.find_all_matches(text, query))
         )
@@ -243,13 +245,14 @@ class SnippetGenerator:
     @staticmethod
     def generate(
         text: str,
-        query: str,
+        query: SearchQuery,
         max_length: NonNegativeInt = DEFAULT_SNIPPET_MAX_LENGTH,
         max_snippets: NonNegativeInt = DEFAULT_MAX_SNIPPETS,
     ) -> list[str]:
         """Generate snippets from text."""
-        if not text or not query:
-            logger.warning("Empty text or query for generate_snippets")
+        assert query is not None
+        if not text:
+            logger.warning("Empty text for generate_snippets")
             return []
 
         candidates = (
@@ -270,7 +273,7 @@ class SnippetGenerator:
     @staticmethod
     def from_summary(
         summary: str,
-        query: str,
+        query: SearchQuery,
         max_snippets: NonNegativeInt = LONG_SUMMARY_MAX_SNIPPETS,
     ) -> list[str]:
         """Generate snippets from summary text."""
@@ -278,9 +281,9 @@ class SnippetGenerator:
 
     @staticmethod
     def combine_sources(
-        summary: str | None,
+        summary: NonEmptyString | None,
         webvtt: WebVTTContent | None,
-        query: str,
+        query: SearchQuery,
         max_total: NonNegativeInt = DEFAULT_MAX_SNIPPETS,
     ) -> tuple[list[str], NonNegativeInt]:
         """Combine snippets from multiple sources and return total match count.
@@ -289,6 +292,11 @@ class SnippetGenerator:
 
         snippets can be empty for real in case of e.g. title match
         """
+
+        assert (
+            summary is not None or webvtt is not None
+        ), "At least one source must be present"
+
         webvtt_matches = 0
         summary_matches = 0
 
@@ -355,8 +363,8 @@ class SearchController:
                 else_=rooms.c.name,
             ).label("room_name"),
         ]
-
-        if params.query_text:
+        search_query = None
+        if params.query_text is not None:
             search_query = sqlalchemy.func.websearch_to_tsquery(
                 "english", params.query_text
             )
@@ -373,7 +381,9 @@ class SearchController:
             transcripts.join(rooms, transcripts.c.room_id == rooms.c.id, isouter=True)
         )
 
-        if params.query_text:
+        if params.query_text is not None:
+            # because already initialized based on params.query_text presence above
+            assert search_query is not None
             base_query = base_query.where(
                 transcripts.c.search_vector_en.op("@@")(search_query)
             )
@@ -393,7 +403,7 @@ class SearchController:
                 transcripts.c.source_kind == params.source_kind
             )
 
-        if params.query_text:
+        if params.query_text is not None:
             order_by = sqlalchemy.desc(sqlalchemy.text("rank"))
         else:
             order_by = sqlalchemy.desc(transcripts.c.created_at)
@@ -407,19 +417,29 @@ class SearchController:
         )
         total = await get_database().fetch_val(count_query)
 
-        def _process_result(r) -> SearchResult:
+        def _process_result(r: DbRecord) -> SearchResult:
             r_dict: Dict[str, Any] = dict(r)
+
             webvtt_raw: str | None = r_dict.pop("webvtt", None)
+            webvtt: WebVTTContent | None
             if webvtt_raw:
                 webvtt = WebVTTProcessor.parse(webvtt_raw)
             else:
                 webvtt = None
-            long_summary: str | None = r_dict.pop("long_summary", None)
+
+            long_summary_r: str | None = r_dict.pop("long_summary", None)
+            long_summary: NonEmptyString = try_parse_non_empty_string(long_summary_r)
             room_name: str | None = r_dict.pop("room_name", None)
             db_result = SearchResultDB.model_validate(r_dict)
 
-            snippets, total_match_count = SnippetGenerator.combine_sources(
-                long_summary, webvtt, params.query_text, DEFAULT_MAX_SNIPPETS
+            at_least_one_source = webvtt is not None or long_summary is not None
+            has_query = params.query_text is not None
+            snippets, total_match_count = (
+                SnippetGenerator.combine_sources(
+                    long_summary, webvtt, params.query_text, DEFAULT_MAX_SNIPPETS
+                )
+                if has_query and at_least_one_source
+                else ([], 0)
             )
 
             return SearchResult(

server/reflector/db/transcripts.py

@@ -122,6 +122,15 @@ def generate_transcript_name() -> str:
     return f"Transcript {now.strftime('%Y-%m-%d %H:%M:%S')}"
 
 
+TranscriptStatus = Literal[
+    "idle", "uploaded", "recording", "processing", "error", "ended"
+]
+
+
+class StrValue(BaseModel):
+    value: str
+
+
 class AudioWaveform(BaseModel):
     data: list[float]
 
@@ -185,7 +194,7 @@ class Transcript(BaseModel):
     id: str = Field(default_factory=generate_uuid4)
     user_id: str | None = None
     name: str = Field(default_factory=generate_transcript_name)
-    status: str = "idle"
+    status: TranscriptStatus = "idle"
     duration: float = 0
     created_at: datetime = Field(default_factory=lambda: datetime.now(timezone.utc))
     title: str | None = None
@@ -638,6 +647,19 @@ class TranscriptController:
         query = transcripts.delete().where(transcripts.c.recording_id == recording_id)
         await get_database().execute(query)
 
+    @staticmethod
+    def user_can_mutate(transcript: Transcript, user_id: str | None) -> bool:
+        """
+        Returns True if the given user is allowed to modify the transcript.
+
+        Policy:
+        - Anonymous transcripts (user_id is None) cannot be modified via API
+        - Only the owner (matching user_id) can modify their transcript
+        """
+        if transcript.user_id is None:
+            return False
+        return user_id and transcript.user_id == user_id
+
     @asynccontextmanager
     async def transaction(self):
         """
@@ -732,5 +754,27 @@ class TranscriptController:
         transcript.delete_participant(participant_id)
         await self.update(transcript, {"participants": transcript.participants_dump()})
 
+    async def set_status(
+        self, transcript_id: str, status: TranscriptStatus
+    ) -> TranscriptEvent | None:
+        """
+        Update the status of a transcript
+
+        Will add an event STATUS + update the status field of transcript
+        """
+        async with self.transaction():
+            transcript = await self.get_by_id(transcript_id)
+            if not transcript:
+                raise Exception(f"Transcript {transcript_id} not found")
+            if transcript.status == status:
+                return
+            resp = await self.append_event(
+                transcript=transcript,
+                event="STATUS",
+                data=StrValue(value=status),
+            )
+            await self.update(transcript, {"status": status})
+        return resp
+
 
 transcripts_controller = TranscriptController()

server/reflector/pipelines/MULTITRACK_FIX_SUMMARY.md

@@ -0,0 +1,84 @@
+# Multitrack Pipeline Fix Summary
+
+## Problem
+Whisper timestamps were incorrect because it ignores leading silence in audio files. Daily.co tracks can have arbitrary amounts of silence before speech starts.
+
+## Solution
+**Pad tracks BEFORE transcription using stream metadata `start_time`**
+
+This makes Whisper timestamps automatically correct relative to recording start.
+
+## Key Changes in `main_multitrack_pipeline_fixed.py`
+
+### 1. Added `pad_track_for_transcription()` method (lines 55-172)
+
+```python
+async def pad_track_for_transcription(
+    self,
+    track_data: bytes,
+    track_idx: int,
+    storage,
+) -> tuple[bytes, str]:
+```
+
+- Extracts stream metadata `start_time` using PyAV
+- Creates PyAV filter graph with `adelay` filter to add padding
+- Stores padded track to S3 and returns URL
+- Uses same audio processing library (PyAV) already in the pipeline
+
+### 2. Modified `process()` method
+
+#### REMOVED (lines 255-302):
+- Entire filename parsing for offsets - NOT NEEDED ANYMORE
+- The complex regex parsing of Daily.co filenames
+- Offset adjustment after transcription
+
+#### ADDED (lines 371-382):
+- Padding step BEFORE transcription:
+```python
+# PAD TRACKS BEFORE TRANSCRIPTION - THIS IS THE KEY FIX!
+padded_track_urls: list[str] = []
+for idx, data in enumerate(track_datas):
+    if not data:
+        padded_track_urls.append("")
+        continue
+
+    _, padded_url = await self.pad_track_for_transcription(
+        data, idx, storage
+    )
+    padded_track_urls.append(padded_url)
+```
+
+#### MODIFIED (lines 385-435):
+- Transcribe PADDED tracks instead of raw tracks
+- Removed all timestamp offset adjustment code
+- Just set speaker ID - timestamps already correct!
+
+```python
+# NO OFFSET ADJUSTMENT NEEDED!
+# Timestamps are already correct because we transcribed padded tracks
+# Just set speaker ID
+for w in t.words:
+    w.speaker = idx
+```
+
+## Why This Works
+
+1. **Stream metadata is authoritative**: Daily.co sets `start_time` in the WebM container
+2. **PyAV respects metadata**: `audio_stream.start_time * audio_stream.time_base` gives seconds
+3. **Padding before transcription**: Whisper sees continuous audio from time 0
+4. **Automatic alignment**: Word at 51s in padded track = 51s in recording
+
+## Testing
+
+Process the test recording (daily-20251020193458) and verify:
+- Participant 0 words appear at ~2s
+- Participant 1 words appear at ~51s
+- No word interleaving
+- Correct chronological order
+
+## Files
+
+- **Original**: `main_multitrack_pipeline.py`
+- **Fixed**: `main_multitrack_pipeline_fixed.py`
+- **Test data**: `/Users/firfi/work/clients/monadical/reflector/1760988935484-*.webm`

server/reflector/pipelines/main_file_pipeline.py

@@ -7,18 +7,28 @@ Uses parallel processing for transcription, diarization, and waveform generation
 """
 
 import asyncio
+import uuid
 from pathlib import Path
 
 import av
 import structlog
-from celery import shared_task
+from celery import chain, shared_task
 
+from reflector.asynctask import asynctask
+from reflector.db.rooms import rooms_controller
 from reflector.db.transcripts import (
+    SourceKind,
     Transcript,
+    TranscriptStatus,
     transcripts_controller,
 )
 from reflector.logger import logger
-from reflector.pipelines.main_live_pipeline import PipelineMainBase, asynctask
+from reflector.pipelines.main_live_pipeline import (
+    PipelineMainBase,
+    broadcast_to_sockets,
+    task_cleanup_consent,
+    task_pipeline_post_to_zulip,
+)
 from reflector.processors import (
     AudioFileWriterProcessor,
     TranscriptFinalSummaryProcessor,
@@ -43,6 +53,7 @@ from reflector.processors.types import (
 )
 from reflector.settings import settings
 from reflector.storage import get_transcripts_storage
+from reflector.worker.webhook import send_transcript_webhook
 
 
 class EmptyPipeline:
@@ -83,12 +94,27 @@ class PipelineMainFile(PipelineMainBase):
                 exc_info=result,
             )
 
+    @broadcast_to_sockets
+    async def set_status(self, transcript_id: str, status: TranscriptStatus):
+        async with self.lock_transaction():
+            return await transcripts_controller.set_status(transcript_id, status)
+
     async def process(self, file_path: Path):
         """Main entry point for file processing"""
         self.logger.info(f"Starting file pipeline for {file_path}")
 
         transcript = await self.get_transcript()
 
+        # Clear transcript as we're going to regenerate everything
+        async with self.transaction():
+            await transcripts_controller.update(
+                transcript,
+                {
+                    "events": [],
+                    "topics": [],
+                },
+            )
+
         # Extract audio and write to transcript location
         audio_path = await self.extract_and_write_audio(file_path, transcript)
 
@@ -105,6 +131,8 @@ class PipelineMainFile(PipelineMainBase):
 
         self.logger.info("File pipeline complete")
 
+        await self.set_status(transcript.id, "ended")
+
     async def extract_and_write_audio(
         self, file_path: Path, transcript: Transcript
     ) -> Path:
@@ -353,6 +381,28 @@ class PipelineMainFile(PipelineMainBase):
         await processor.flush()
 
 
+@shared_task
+@asynctask
+async def task_send_webhook_if_needed(*, transcript_id: str):
+    """Send webhook if this is a room recording with webhook configured"""
+    transcript = await transcripts_controller.get_by_id(transcript_id)
+    if not transcript:
+        return
+
+    if transcript.source_kind == SourceKind.ROOM and transcript.room_id:
+        room = await rooms_controller.get_by_id(transcript.room_id)
+        if room and room.webhook_url:
+            logger.info(
+                "Dispatching webhook",
+                transcript_id=transcript_id,
+                room_id=room.id,
+                webhook_url=room.webhook_url,
+            )
+            send_transcript_webhook.delay(
+                transcript_id, room.id, event_id=uuid.uuid4().hex
+            )
+
+
 @shared_task
 @asynctask
 async def task_pipeline_file_process(*, transcript_id: str):
@@ -362,14 +412,28 @@ async def task_pipeline_file_process(*, transcript_id: str):
     if not transcript:
         raise Exception(f"Transcript {transcript_id} not found")
 
-    # Find the file to process
-    audio_file = next(transcript.data_path.glob("upload.*"), None)
-    if not audio_file:
-        audio_file = next(transcript.data_path.glob("audio.*"), None)
-
-    if not audio_file:
-        raise Exception("No audio file found to process")
-
-    # Run file pipeline
     pipeline = PipelineMainFile(transcript_id=transcript_id)
-    await pipeline.process(audio_file)
+    try:
+        await pipeline.set_status(transcript_id, "processing")
+
+        # Find the file to process
+        audio_file = next(transcript.data_path.glob("upload.*"), None)
+        if not audio_file:
+            audio_file = next(transcript.data_path.glob("audio.*"), None)
+
+        if not audio_file:
+            raise Exception("No audio file found to process")
+
+        await pipeline.process(audio_file)
+
+    except Exception:
+        await pipeline.set_status(transcript_id, "error")
+        raise
+
+    # Run post-processing chain: consent cleanup -> zulip -> webhook
+    post_chain = chain(
+        task_cleanup_consent.si(transcript_id=transcript_id),
+        task_pipeline_post_to_zulip.si(transcript_id=transcript_id),
+        task_send_webhook_if_needed.si(transcript_id=transcript_id),
+    )
+    post_chain.delay()

server/reflector/pipelines/main_live_pipeline.py

@@ -22,7 +22,7 @@ from celery import chord, current_task, group, shared_task
 from pydantic import BaseModel
 from structlog import BoundLogger as Logger
 
-from reflector.db import get_database
+from reflector.asynctask import asynctask
 from reflector.db.meetings import meeting_consent_controller, meetings_controller
 from reflector.db.recordings import recordings_controller
 from reflector.db.rooms import rooms_controller
@@ -32,6 +32,7 @@ from reflector.db.transcripts import (
     TranscriptFinalLongSummary,
     TranscriptFinalShortSummary,
     TranscriptFinalTitle,
+    TranscriptStatus,
     TranscriptText,
     TranscriptTopic,
     TranscriptWaveform,
@@ -40,8 +41,9 @@ from reflector.db.transcripts import (
 from reflector.logger import logger
 from reflector.pipelines.runner import PipelineMessage, PipelineRunner
 from reflector.processors import (
-    AudioChunkerProcessor,
+    AudioChunkerAutoProcessor,
     AudioDiarizationAutoProcessor,
+    AudioDownscaleProcessor,
     AudioFileWriterProcessor,
     AudioMergeProcessor,
     AudioTranscriptAutoProcessor,
@@ -68,29 +70,6 @@ from reflector.zulip import (
 )
 
 
-def asynctask(f):
-    @functools.wraps(f)
-    def wrapper(*args, **kwargs):
-        async def run_with_db():
-            database = get_database()
-            await database.connect()
-            try:
-                return await f(*args, **kwargs)
-            finally:
-                await database.disconnect()
-
-        coro = run_with_db()
-        try:
-            loop = asyncio.get_running_loop()
-        except RuntimeError:
-            loop = None
-        if loop and loop.is_running():
-            return loop.run_until_complete(coro)
-        return asyncio.run(coro)
-
-    return wrapper
-
-
 def broadcast_to_sockets(func):
     """
     Decorator to broadcast transcript event to websockets
@@ -106,6 +85,20 @@ def broadcast_to_sockets(func):
             message=resp.model_dump(mode="json"),
         )
 
+        transcript = await transcripts_controller.get_by_id(self.transcript_id)
+        if transcript and transcript.user_id:
+            # Emit only relevant events to the user room to avoid noisy updates.
+            # Allowed: STATUS, FINAL_TITLE, DURATION. All are prefixed with TRANSCRIPT_
+            allowed_user_events = {"STATUS", "FINAL_TITLE", "DURATION"}
+            if resp.event in allowed_user_events:
+                await self.ws_manager.send_json(
+                    room_id=f"user:{transcript.user_id}",
+                    message={
+                        "event": f"TRANSCRIPT_{resp.event}",
+                        "data": {"id": self.transcript_id, **resp.data},
+                    },
+                )
+
     return wrapper
 
 
@@ -187,8 +180,15 @@ class PipelineMainBase(PipelineRunner[PipelineMessage], Generic[PipelineMessage]
         ]
 
     @asynccontextmanager
-    async def transaction(self):
+    async def lock_transaction(self):
+        # This lock is to prevent multiple processor starting adding
+        # into event array at the same time
         async with self._lock:
+            yield
+
+    @asynccontextmanager
+    async def transaction(self):
+        async with self.lock_transaction():
             async with transcripts_controller.transaction():
                 yield
 
@@ -197,14 +197,14 @@ class PipelineMainBase(PipelineRunner[PipelineMessage], Generic[PipelineMessage]
         # if it's the first part, update the status of the transcript
         # but do not set the ended status yet.
         if isinstance(self, PipelineMainLive):
-            status_mapping = {
+            status_mapping: dict[str, TranscriptStatus] = {
                 "started": "recording",
                 "push": "recording",
                 "flush": "processing",
                 "error": "error",
             }
         elif isinstance(self, PipelineMainFinalSummaries):
-            status_mapping = {
+            status_mapping: dict[str, TranscriptStatus] = {
                 "push": "processing",
                 "flush": "processing",
                 "error": "error",
@@ -220,22 +220,8 @@ class PipelineMainBase(PipelineRunner[PipelineMessage], Generic[PipelineMessage]
             return
 
         # when the status of the pipeline changes, update the transcript
-        async with self.transaction():
-            transcript = await self.get_transcript()
-            if status == transcript.status:
-                return
-            resp = await transcripts_controller.append_event(
-                transcript=transcript,
-                event="STATUS",
-                data=StrValue(value=status),
-            )
-            await transcripts_controller.update(
-                transcript,
-                {
-                    "status": status,
-                },
-            )
-            return resp
+        async with self._lock:
+            return await transcripts_controller.set_status(self.transcript_id, status)
 
     @broadcast_to_sockets
     async def on_transcript(self, data):
@@ -365,7 +351,8 @@ class PipelineMainLive(PipelineMainBase):
                 path=transcript.audio_wav_filename,
                 on_duration=self.on_duration,
             ),
-            AudioChunkerProcessor(),
+            AudioDownscaleProcessor(),
+            AudioChunkerAutoProcessor(),
             AudioMergeProcessor(),
             AudioTranscriptAutoProcessor.as_threaded(),
             TranscriptLinerProcessor(),
@@ -792,7 +779,7 @@ def pipeline_post(*, transcript_id: str):
         chain_final_summaries,
     ) | task_pipeline_post_to_zulip.si(transcript_id=transcript_id)
 
-    chain.delay()
+    return chain.delay()
 
 
 @get_transcript

server/reflector/pipelines/main_multitrack_pipeline.backup.py

@@ -0,0 +1,510 @@
+import asyncio
+import io
+from fractions import Fraction
+
+import av
+import boto3
+import structlog
+from av.audio.resampler import AudioResampler
+from celery import chain, shared_task
+
+from reflector.asynctask import asynctask
+from reflector.db.transcripts import (
+    TranscriptStatus,
+    TranscriptText,
+    transcripts_controller,
+)
+from reflector.logger import logger
+from reflector.pipelines.main_file_pipeline import task_send_webhook_if_needed
+from reflector.pipelines.main_live_pipeline import (
+    PipelineMainBase,
+    task_cleanup_consent,
+    task_pipeline_post_to_zulip,
+)
+from reflector.processors import (
+    AudioFileWriterProcessor,
+    TranscriptFinalSummaryProcessor,
+    TranscriptFinalTitleProcessor,
+    TranscriptTopicDetectorProcessor,
+)
+from reflector.processors.file_transcript import FileTranscriptInput
+from reflector.processors.file_transcript_auto import FileTranscriptAutoProcessor
+from reflector.processors.types import TitleSummary
+from reflector.processors.types import (
+    Transcript as TranscriptType,
+)
+from reflector.settings import settings
+from reflector.storage import get_transcripts_storage
+
+
+class EmptyPipeline:
+    def __init__(self, logger: structlog.BoundLogger):
+        self.logger = logger
+
+    def get_pref(self, k, d=None):
+        return d
+
+    async def emit(self, event):
+        pass
+
+
+class PipelineMainMultitrack(PipelineMainBase):
+    """Process multiple participant tracks for a transcript without mixing audio."""
+
+    def __init__(self, transcript_id: str):
+        super().__init__(transcript_id=transcript_id)
+        self.logger = logger.bind(transcript_id=self.transcript_id)
+        self.empty_pipeline = EmptyPipeline(logger=self.logger)
+
+    async def mixdown_tracks(
+        self,
+        track_datas: list[bytes],
+        writer: AudioFileWriterProcessor,
+        offsets_seconds: list[float] | None = None,
+    ) -> None:
+        """
+        Minimal multi-track mixdown using a PyAV filter graph (amix), no resampling.
+        """
+
+        # Discover target sample rate from first decodable frame
+        target_sample_rate: int | None = None
+        for data in track_datas:
+            if not data:
+                continue
+            try:
+                container = av.open(io.BytesIO(data))
+                try:
+                    for frame in container.decode(audio=0):
+                        target_sample_rate = frame.sample_rate
+                        break
+                finally:
+                    container.close()
+            except Exception:
+                continue
+            if target_sample_rate:
+                break
+
+        if not target_sample_rate:
+            self.logger.warning("Mixdown skipped - no decodable audio frames found")
+            return
+
+        # Build PyAV filter graph:
+        # N abuffer (s32/stereo)
+        #   -> optional adelay per input (for alignment)
+        #   -> amix (s32)
+        #   -> aformat(s16)
+        #   -> sink
+        graph = av.filter.Graph()
+        inputs = []
+        valid_track_datas = [d for d in track_datas if d]
+        # Align offsets list with the filtered inputs (skip empties)
+        input_offsets_seconds = None
+        if offsets_seconds is not None:
+            input_offsets_seconds = [
+                offsets_seconds[i] for i, d in enumerate(track_datas) if d
+            ]
+        for idx, data in enumerate(valid_track_datas):
+            args = (
+                f"time_base=1/{target_sample_rate}:"
+                f"sample_rate={target_sample_rate}:"
+                f"sample_fmt=s32:"
+                f"channel_layout=stereo"
+            )
+            in_ctx = graph.add("abuffer", args=args, name=f"in{idx}")
+            inputs.append(in_ctx)
+
+        if not inputs:
+            self.logger.warning("Mixdown skipped - no valid inputs for graph")
+            return
+
+        mixer = graph.add("amix", args=f"inputs={len(inputs)}:normalize=0", name="mix")
+
+        fmt = graph.add(
+            "aformat",
+            args=(
+                f"sample_fmts=s32:channel_layouts=stereo:sample_rates={target_sample_rate}"
+            ),
+            name="fmt",
+        )
+
+        sink = graph.add("abuffersink", name="out")
+
+        # Optional per-input delay before mixing
+        delays_ms: list[int] = []
+        if input_offsets_seconds is not None:
+            base = min(input_offsets_seconds) if input_offsets_seconds else 0.0
+            delays_ms = [
+                max(0, int(round((o - base) * 1000))) for o in input_offsets_seconds
+            ]
+        else:
+            delays_ms = [0 for _ in inputs]
+
+        for idx, in_ctx in enumerate(inputs):
+            delay_ms = delays_ms[idx] if idx < len(delays_ms) else 0
+            if delay_ms > 0:
+                # adelay requires one value per channel; use same for stereo
+                adelay = graph.add(
+                    "adelay",
+                    args=f"delays={delay_ms}|{delay_ms}:all=1",
+                    name=f"delay{idx}",
+                )
+                in_ctx.link_to(adelay)
+                adelay.link_to(mixer, 0, idx)
+            else:
+                in_ctx.link_to(mixer, 0, idx)
+        mixer.link_to(fmt)
+        fmt.link_to(sink)
+        graph.configure()
+
+        # Open containers for decoding
+        containers = []
+        for i, d in enumerate(valid_track_datas):
+            try:
+                c = av.open(io.BytesIO(d))
+                containers.append(c)
+            except Exception as e:
+                self.logger.warning(
+                    "Mixdown: failed to open container", input=i, error=str(e)
+                )
+                containers.append(None)
+        # Filter out Nones for decoders
+        containers = [c for c in containers if c is not None]
+        decoders = [c.decode(audio=0) for c in containers]
+        active = [True] * len(decoders)
+        # Per-input resamplers to enforce s32/stereo at the same rate (no resample of rate)
+        resamplers = [
+            AudioResampler(format="s32", layout="stereo", rate=target_sample_rate)
+            for _ in decoders
+        ]
+
+        try:
+            # Round-robin feed frames into graph, pull mixed frames as they become available
+            while any(active):
+                for i, (dec, is_active) in enumerate(zip(decoders, active)):
+                    if not is_active:
+                        continue
+                    try:
+                        frame = next(dec)
+                    except StopIteration:
+                        active[i] = False
+                        continue
+
+                    # Enforce same sample rate; convert format/layout to s16/stereo (no resample)
+                    if frame.sample_rate != target_sample_rate:
+                        # Skip frames with differing rate
+                        continue
+                    out_frames = resamplers[i].resample(frame) or []
+                    for rf in out_frames:
+                        rf.sample_rate = target_sample_rate
+                        rf.time_base = Fraction(1, target_sample_rate)
+                        inputs[i].push(rf)
+
+                    # Drain available mixed frames
+                    while True:
+                        try:
+                            mixed = sink.pull()
+                        except Exception:
+                            break
+                        mixed.sample_rate = target_sample_rate
+                        mixed.time_base = Fraction(1, target_sample_rate)
+                        await writer.push(mixed)
+
+            # Signal EOF to inputs and drain remaining
+            for in_ctx in inputs:
+                in_ctx.push(None)
+            while True:
+                try:
+                    mixed = sink.pull()
+                except Exception:
+                    break
+                mixed.sample_rate = target_sample_rate
+                mixed.time_base = Fraction(1, target_sample_rate)
+                await writer.push(mixed)
+        finally:
+            for c in containers:
+                c.close()
+
+    async def set_status(self, transcript_id: str, status: TranscriptStatus):
+        async with self.lock_transaction():
+            return await transcripts_controller.set_status(transcript_id, status)
+
+    async def process(self, bucket_name: str, track_keys: list[str]):
+        transcript = await self.get_transcript()
+
+        s3 = boto3.client(
+            "s3",
+            region_name=settings.RECORDING_STORAGE_AWS_REGION,
+            aws_access_key_id=settings.RECORDING_STORAGE_AWS_ACCESS_KEY_ID,
+            aws_secret_access_key=settings.RECORDING_STORAGE_AWS_SECRET_ACCESS_KEY,
+        )
+
+        storage = get_transcripts_storage()
+
+        # Pre-download bytes for all tracks for mixing and transcription
+        track_datas: list[bytes] = []
+        for key in track_keys:
+            try:
+                obj = s3.get_object(Bucket=bucket_name, Key=key)
+                track_datas.append(obj["Body"].read())
+            except Exception as e:
+                self.logger.warning(
+                    "Skipping track - cannot read S3 object", key=key, error=str(e)
+                )
+                track_datas.append(b"")
+
+        # Extract offsets from Daily.co filename timestamps
+        # Format: {rec_start_ts}-{uuid}-{media_type}-{track_start_ts}.{ext}
+        # Example: 1760988935484-uuid-cam-audio-1760988935922
+        import re
+
+        offsets_seconds: list[float] = []
+        recording_start_ts: int | None = None
+
+        for key in track_keys:
+            # Parse Daily.co raw-tracks filename pattern
+            match = re.search(r"(\d+)-([0-9a-f-]{36})-(cam-audio)-(\d+)", key)
+            if not match:
+                self.logger.warning(
+                    "Track key doesn't match Daily.co pattern, using 0.0 offset",
+                    key=key,
+                )
+                offsets_seconds.append(0.0)
+                continue
+
+            rec_start_ts = int(match.group(1))
+            track_start_ts = int(match.group(4))
+
+            # Validate all tracks belong to same recording
+            if recording_start_ts is None:
+                recording_start_ts = rec_start_ts
+            elif rec_start_ts != recording_start_ts:
+                self.logger.error(
+                    "Track belongs to different recording",
+                    key=key,
+                    expected_start=recording_start_ts,
+                    got_start=rec_start_ts,
+                )
+                offsets_seconds.append(0.0)
+                continue
+
+            # Calculate offset in seconds
+            offset_ms = track_start_ts - rec_start_ts
+            offset_s = offset_ms / 1000.0
+
+            self.logger.info(
+                "Parsed track offset from filename",
+                key=key,
+                recording_start=rec_start_ts,
+                track_start=track_start_ts,
+                offset_seconds=offset_s,
+            )
+
+            offsets_seconds.append(max(0.0, offset_s))
+
+        # Mixdown all available tracks into transcript.audio_mp3_filename, preserving sample rate
+        try:
+            mp3_writer = AudioFileWriterProcessor(
+                path=str(transcript.audio_mp3_filename)
+            )
+            await self.mixdown_tracks(track_datas, mp3_writer, offsets_seconds)
+            await mp3_writer.flush()
+        except Exception as e:
+            self.logger.error("Mixdown failed", error=str(e))
+
+        speaker_transcripts: list[TranscriptType] = []
+        for idx, key in enumerate(track_keys):
+            ext = ".mp4"
+
+            try:
+                obj = s3.get_object(Bucket=bucket_name, Key=key)
+                data = obj["Body"].read()
+            except Exception as e:
+                self.logger.error(
+                    "Skipping track - cannot read S3 object", key=key, error=str(e)
+                )
+                continue
+
+            storage_path = f"file_pipeline/{transcript.id}/tracks/track_{idx}{ext}"
+            try:
+                await storage.put_file(storage_path, data)
+                audio_url = await storage.get_file_url(storage_path)
+            except Exception as e:
+                self.logger.error(
+                    "Skipping track - cannot upload to storage", key=key, error=str(e)
+                )
+                continue
+
+            try:
+                t = await self.transcribe_file(audio_url, transcript.source_language)
+            except Exception as e:
+                self.logger.error(
+                    "Transcription via default backend failed, trying local whisper",
+                    key=key,
+                    url=audio_url,
+                    error=str(e),
+                )
+                try:
+                    fallback = FileTranscriptAutoProcessor(name="whisper")
+                    result = None
+
+                    async def capture_result(r):
+                        nonlocal result
+                        result = r
+
+                    fallback.on(capture_result)
+                    await fallback.push(
+                        FileTranscriptInput(
+                            audio_url=audio_url, language=transcript.source_language
+                        )
+                    )
+                    await fallback.flush()
+                    if not result:
+                        raise Exception("No transcript captured in fallback")
+                    t = result
+                except Exception as e2:
+                    self.logger.error(
+                        "Skipping track - transcription failed after fallback",
+                        key=key,
+                        url=audio_url,
+                        error=str(e2),
+                    )
+                    continue
+
+            if not t.words:
+                continue
+            # Shift word timestamps by the track's offset so all are relative to 00:00
+            track_offset = offsets_seconds[idx] if idx < len(offsets_seconds) else 0.0
+            for w in t.words:
+                try:
+                    if hasattr(w, "start") and w.start is not None:
+                        w.start = float(w.start) + track_offset
+                    if hasattr(w, "end") and w.end is not None:
+                        w.end = float(w.end) + track_offset
+                except Exception:
+                    pass
+                w.speaker = idx
+            speaker_transcripts.append(t)
+
+        if not speaker_transcripts:
+            raise Exception("No valid track transcriptions")
+
+        merged_words = []
+        for t in speaker_transcripts:
+            merged_words.extend(t.words)
+        merged_words.sort(key=lambda w: w.start)
+
+        merged_transcript = TranscriptType(words=merged_words, translation=None)
+
+        await transcripts_controller.append_event(
+            transcript,
+            event="TRANSCRIPT",
+            data=TranscriptText(
+                text=merged_transcript.text, translation=merged_transcript.translation
+            ),
+        )
+
+        topics = await self.detect_topics(merged_transcript, transcript.target_language)
+        await asyncio.gather(
+            self.generate_title(topics),
+            self.generate_summaries(topics),
+            return_exceptions=False,
+        )
+
+        await self.set_status(transcript.id, "ended")
+
+    async def transcribe_file(self, audio_url: str, language: str) -> TranscriptType:
+        processor = FileTranscriptAutoProcessor()
+        input_data = FileTranscriptInput(audio_url=audio_url, language=language)
+
+        result: TranscriptType | None = None
+
+        async def capture_result(transcript):
+            nonlocal result
+            result = transcript
+
+        processor.on(capture_result)
+        await processor.push(input_data)
+        await processor.flush()
+
+        if not result:
+            raise ValueError("No transcript captured")
+
+        return result
+
+    async def detect_topics(
+        self, transcript: TranscriptType, target_language: str
+    ) -> list[TitleSummary]:
+        chunk_size = 300
+        topics: list[TitleSummary] = []
+
+        async def on_topic(topic: TitleSummary):
+            topics.append(topic)
+            return await self.on_topic(topic)
+
+        topic_detector = TranscriptTopicDetectorProcessor(callback=on_topic)
+        topic_detector.set_pipeline(self.empty_pipeline)
+
+        for i in range(0, len(transcript.words), chunk_size):
+            chunk_words = transcript.words[i : i + chunk_size]
+            if not chunk_words:
+                continue
+
+            chunk_transcript = TranscriptType(
+                words=chunk_words, translation=transcript.translation
+            )
+            await topic_detector.push(chunk_transcript)
+
+        await topic_detector.flush()
+        return topics
+
+    async def generate_title(self, topics: list[TitleSummary]):
+        if not topics:
+            self.logger.warning("No topics for title generation")
+            return
+
+        processor = TranscriptFinalTitleProcessor(callback=self.on_title)
+        processor.set_pipeline(self.empty_pipeline)
+
+        for topic in topics:
+            await processor.push(topic)
+
+        await processor.flush()
+
+    async def generate_summaries(self, topics: list[TitleSummary]):
+        if not topics:
+            self.logger.warning("No topics for summary generation")
+            return
+
+        transcript = await self.get_transcript()
+        processor = TranscriptFinalSummaryProcessor(
+            transcript=transcript,
+            callback=self.on_long_summary,
+            on_short_summary=self.on_short_summary,
+        )
+        processor.set_pipeline(self.empty_pipeline)
+
+        for topic in topics:
+            await processor.push(topic)
+
+        await processor.flush()
+
+
+@shared_task
+@asynctask
+async def task_pipeline_multitrack_process(
+    *, transcript_id: str, bucket_name: str, track_keys: list[str]
+):
+    pipeline = PipelineMainMultitrack(transcript_id=transcript_id)
+    try:
+        await pipeline.set_status(transcript_id, "processing")
+        await pipeline.process(bucket_name, track_keys)
+    except Exception:
+        await pipeline.set_status(transcript_id, "error")
+        raise
+
+    post_chain = chain(
+        task_cleanup_consent.si(transcript_id=transcript_id),
+        task_pipeline_post_to_zulip.si(transcript_id=transcript_id),
+        task_send_webhook_if_needed.si(transcript_id=transcript_id),
+    )
+    post_chain.delay()

server/reflector/pipelines/main_multitrack_pipeline.py

@@ -0,0 +1,654 @@
+import asyncio
+import io
+from fractions import Fraction
+
+import av
+import boto3
+import structlog
+from av.audio.resampler import AudioResampler
+from celery import chain, shared_task
+
+from reflector.asynctask import asynctask
+from reflector.db.transcripts import (
+    TranscriptStatus,
+    TranscriptWaveform,
+    transcripts_controller,
+)
+from reflector.logger import logger
+from reflector.pipelines.main_file_pipeline import task_send_webhook_if_needed
+from reflector.pipelines.main_live_pipeline import (
+    PipelineMainBase,
+    broadcast_to_sockets,
+    task_cleanup_consent,
+    task_pipeline_post_to_zulip,
+)
+from reflector.processors import (
+    AudioFileWriterProcessor,
+    TranscriptFinalSummaryProcessor,
+    TranscriptFinalTitleProcessor,
+    TranscriptTopicDetectorProcessor,
+)
+from reflector.processors.audio_waveform_processor import AudioWaveformProcessor
+from reflector.processors.file_transcript import FileTranscriptInput
+from reflector.processors.file_transcript_auto import FileTranscriptAutoProcessor
+from reflector.processors.types import TitleSummary
+from reflector.processors.types import (
+    Transcript as TranscriptType,
+)
+from reflector.settings import settings
+from reflector.storage import get_transcripts_storage
+
+
+class EmptyPipeline:
+    def __init__(self, logger: structlog.BoundLogger):
+        self.logger = logger
+
+    def get_pref(self, k, d=None):
+        return d
+
+    async def emit(self, event):
+        pass
+
+
+class PipelineMainMultitrack(PipelineMainBase):
+    """Process multiple participant tracks for a transcript without mixing audio."""
+
+    def __init__(self, transcript_id: str):
+        super().__init__(transcript_id=transcript_id)
+        self.logger = logger.bind(transcript_id=self.transcript_id)
+        self.empty_pipeline = EmptyPipeline(logger=self.logger)
+
+    async def pad_track_for_transcription(
+        self,
+        track_data: bytes,
+        track_idx: int,
+        storage,
+    ) -> tuple[bytes, str]:
+        """
+        Pad a single track with silence based on stream metadata start_time.
+        This ensures Whisper timestamps will be relative to recording start.
+        Uses ffmpeg subprocess approach proven to work with python-raw-tracks-align.
+
+        Returns: (padded_data, storage_url)
+        """
+        import json
+        import math
+        import subprocess
+        import tempfile
+
+        if not track_data:
+            return b"", ""
+
+        transcript = await self.get_transcript()
+
+        # Create temp files for ffmpeg processing
+        with tempfile.NamedTemporaryFile(suffix=".webm", delete=False) as input_file:
+            input_file.write(track_data)
+            input_file_path = input_file.name
+
+        output_file_path = input_file_path.replace(".webm", "_padded.webm")
+
+        try:
+            # Get stream metadata using ffprobe
+            ffprobe_cmd = [
+                "ffprobe",
+                "-v",
+                "error",
+                "-show_entries",
+                "stream=start_time",
+                "-of",
+                "json",
+                input_file_path,
+            ]
+
+            result = subprocess.run(
+                ffprobe_cmd, capture_output=True, text=True, check=True
+            )
+            metadata = json.loads(result.stdout)
+
+            # Extract start_time from stream metadata
+            start_time_seconds = 0.0
+            if metadata.get("streams") and len(metadata["streams"]) > 0:
+                start_time_str = metadata["streams"][0].get("start_time", "0")
+                start_time_seconds = float(start_time_str)
+
+            self.logger.info(
+                f"Track {track_idx} stream metadata: start_time={start_time_seconds:.3f}s",
+                track_idx=track_idx,
+            )
+
+            # If no padding needed, use original
+            if start_time_seconds <= 0:
+                storage_path = f"file_pipeline/{transcript.id}/tracks/original_track_{track_idx}.webm"
+                await storage.put_file(storage_path, track_data)
+                url = await storage.get_file_url(storage_path)
+                return track_data, url
+
+            # Calculate delay in milliseconds
+            delay_ms = math.floor(start_time_seconds * 1000)
+
+            # Run ffmpeg to pad the audio while maintaining WebM/Opus format for Modal compatibility
+            # ffmpeg quirk: aresample needs to come before adelay in the filter chain
+            ffmpeg_cmd = [
+                "ffmpeg",
+                "-hide_banner",
+                "-loglevel",
+                "error",
+                "-y",  # overwrite output
+                "-i",
+                input_file_path,
+                "-af",
+                f"aresample=async=1,adelay={delay_ms}:all=true",
+                "-c:a",
+                "libopus",  # Keep Opus codec for Modal compatibility
+                "-b:a",
+                "128k",  # Standard bitrate for Opus
+                output_file_path,
+            ]
+
+            self.logger.info(
+                f"Padding track {track_idx} with {delay_ms}ms delay using ffmpeg",
+                track_idx=track_idx,
+                delay_ms=delay_ms,
+                command=" ".join(ffmpeg_cmd),
+            )
+
+            result = subprocess.run(ffmpeg_cmd, capture_output=True, text=True)
+            if result.returncode != 0:
+                self.logger.error(
+                    f"ffmpeg padding failed for track {track_idx}",
+                    track_idx=track_idx,
+                    stderr=result.stderr,
+                    returncode=result.returncode,
+                )
+                raise Exception(f"ffmpeg padding failed: {result.stderr}")
+
+            # Read the padded output
+            with open(output_file_path, "rb") as f:
+                padded_data = f.read()
+
+            # Store padded track
+            storage_path = (
+                f"file_pipeline/{transcript.id}/tracks/padded_track_{track_idx}.webm"
+            )
+            await storage.put_file(storage_path, padded_data)
+            padded_url = await storage.get_file_url(storage_path)
+
+            self.logger.info(
+                f"Successfully padded track {track_idx} with {start_time_seconds:.3f}s offset, stored at {storage_path}",
+                track_idx=track_idx,
+                delay_ms=delay_ms,
+                padded_url=padded_url,
+                padded_size=len(padded_data),
+            )
+
+            return padded_data, padded_url
+
+        finally:
+            # Clean up temp files
+            import os
+
+            try:
+                os.unlink(input_file_path)
+            except:
+                pass
+            try:
+                os.unlink(output_file_path)
+            except:
+                pass
+
+    async def mixdown_tracks(
+        self,
+        track_datas: list[bytes],
+        writer: AudioFileWriterProcessor,
+        offsets_seconds: list[float] | None = None,
+    ) -> None:
+        """
+        Minimal multi-track mixdown using a PyAV filter graph (amix), no resampling.
+        """
+
+        # Discover target sample rate from first decodable frame
+        target_sample_rate: int | None = None
+        for data in track_datas:
+            if not data:
+                continue
+            try:
+                container = av.open(io.BytesIO(data))
+                try:
+                    for frame in container.decode(audio=0):
+                        target_sample_rate = frame.sample_rate
+                        break
+                finally:
+                    container.close()
+            except Exception:
+                continue
+            if target_sample_rate:
+                break
+
+        if not target_sample_rate:
+            self.logger.warning("Mixdown skipped - no decodable audio frames found")
+            return
+
+        # Build PyAV filter graph:
+        # N abuffer (s32/stereo)
+        #   -> optional adelay per input (for alignment)
+        #   -> amix (s32)
+        #   -> aformat(s16)
+        #   -> sink
+        graph = av.filter.Graph()
+        inputs = []
+        valid_track_datas = [d for d in track_datas if d]
+        # Align offsets list with the filtered inputs (skip empties)
+        input_offsets_seconds = None
+        if offsets_seconds is not None:
+            input_offsets_seconds = [
+                offsets_seconds[i] for i, d in enumerate(track_datas) if d
+            ]
+        for idx, data in enumerate(valid_track_datas):
+            args = (
+                f"time_base=1/{target_sample_rate}:"
+                f"sample_rate={target_sample_rate}:"
+                f"sample_fmt=s32:"
+                f"channel_layout=stereo"
+            )
+            in_ctx = graph.add("abuffer", args=args, name=f"in{idx}")
+            inputs.append(in_ctx)
+
+        if not inputs:
+            self.logger.warning("Mixdown skipped - no valid inputs for graph")
+            return
+
+        mixer = graph.add("amix", args=f"inputs={len(inputs)}:normalize=0", name="mix")
+
+        fmt = graph.add(
+            "aformat",
+            args=(
+                f"sample_fmts=s32:channel_layouts=stereo:sample_rates={target_sample_rate}"
+            ),
+            name="fmt",
+        )
+
+        sink = graph.add("abuffersink", name="out")
+
+        # Optional per-input delay before mixing
+        delays_ms: list[int] = []
+        if input_offsets_seconds is not None:
+            base = min(input_offsets_seconds) if input_offsets_seconds else 0.0
+            delays_ms = [
+                max(0, int(round((o - base) * 1000))) for o in input_offsets_seconds
+            ]
+        else:
+            delays_ms = [0 for _ in inputs]
+
+        for idx, in_ctx in enumerate(inputs):
+            delay_ms = delays_ms[idx] if idx < len(delays_ms) else 0
+            if delay_ms > 0:
+                # adelay requires one value per channel; use same for stereo
+                adelay = graph.add(
+                    "adelay",
+                    args=f"delays={delay_ms}|{delay_ms}:all=1",
+                    name=f"delay{idx}",
+                )
+                in_ctx.link_to(adelay)
+                adelay.link_to(mixer, 0, idx)
+            else:
+                in_ctx.link_to(mixer, 0, idx)
+        mixer.link_to(fmt)
+        fmt.link_to(sink)
+        graph.configure()
+
+        # Open containers for decoding
+        containers = []
+        for i, d in enumerate(valid_track_datas):
+            try:
+                c = av.open(io.BytesIO(d))
+                containers.append(c)
+            except Exception as e:
+                self.logger.warning(
+                    "Mixdown: failed to open container", input=i, error=str(e)
+                )
+                containers.append(None)
+        # Filter out Nones for decoders
+        containers = [c for c in containers if c is not None]
+        decoders = [c.decode(audio=0) for c in containers]
+        active = [True] * len(decoders)
+        # Per-input resamplers to enforce s32/stereo at the same rate (no resample of rate)
+        resamplers = [
+            AudioResampler(format="s32", layout="stereo", rate=target_sample_rate)
+            for _ in decoders
+        ]
+
+        try:
+            # Round-robin feed frames into graph, pull mixed frames as they become available
+            while any(active):
+                for i, (dec, is_active) in enumerate(zip(decoders, active)):
+                    if not is_active:
+                        continue
+                    try:
+                        frame = next(dec)
+                    except StopIteration:
+                        active[i] = False
+                        continue
+
+                    # Enforce same sample rate; convert format/layout to s16/stereo (no resample)
+                    if frame.sample_rate != target_sample_rate:
+                        # Skip frames with differing rate
+                        continue
+                    out_frames = resamplers[i].resample(frame) or []
+                    for rf in out_frames:
+                        rf.sample_rate = target_sample_rate
+                        rf.time_base = Fraction(1, target_sample_rate)
+                        inputs[i].push(rf)
+
+                    # Drain available mixed frames
+                    while True:
+                        try:
+                            mixed = sink.pull()
+                        except Exception:
+                            break
+                        mixed.sample_rate = target_sample_rate
+                        mixed.time_base = Fraction(1, target_sample_rate)
+                        await writer.push(mixed)
+
+            # Signal EOF to inputs and drain remaining
+            for in_ctx in inputs:
+                in_ctx.push(None)
+            while True:
+                try:
+                    mixed = sink.pull()
+                except Exception:
+                    break
+                mixed.sample_rate = target_sample_rate
+                mixed.time_base = Fraction(1, target_sample_rate)
+                await writer.push(mixed)
+        finally:
+            for c in containers:
+                c.close()
+
+    @broadcast_to_sockets
+    async def set_status(self, transcript_id: str, status: TranscriptStatus):
+        async with self.lock_transaction():
+            return await transcripts_controller.set_status(transcript_id, status)
+
+    async def on_waveform(self, data):
+        async with self.transaction():
+            waveform = TranscriptWaveform(waveform=data)
+            transcript = await self.get_transcript()
+            return await transcripts_controller.append_event(
+                transcript=transcript, event="WAVEFORM", data=waveform
+            )
+
+    async def process(self, bucket_name: str, track_keys: list[str]):
+        transcript = await self.get_transcript()
+
+        s3 = boto3.client(
+            "s3",
+            region_name=settings.RECORDING_STORAGE_AWS_REGION,
+            aws_access_key_id=settings.RECORDING_STORAGE_AWS_ACCESS_KEY_ID,
+            aws_secret_access_key=settings.RECORDING_STORAGE_AWS_SECRET_ACCESS_KEY,
+        )
+
+        storage = get_transcripts_storage()
+
+        # Pre-download bytes for all tracks for mixing and transcription
+        track_datas: list[bytes] = []
+        for key in track_keys:
+            try:
+                obj = s3.get_object(Bucket=bucket_name, Key=key)
+                track_datas.append(obj["Body"].read())
+            except Exception as e:
+                self.logger.warning(
+                    "Skipping track - cannot read S3 object", key=key, error=str(e)
+                )
+                track_datas.append(b"")
+
+        # PAD TRACKS FIRST - this creates full-length tracks with correct timeline
+        padded_track_datas: list[bytes] = []
+        padded_track_urls: list[str] = []
+        for idx, data in enumerate(track_datas):
+            if not data:
+                padded_track_datas.append(b"")
+                padded_track_urls.append("")
+                continue
+
+            padded_data, padded_url = await self.pad_track_for_transcription(
+                data, idx, storage
+            )
+            padded_track_datas.append(padded_data)
+            padded_track_urls.append(padded_url)
+            self.logger.info(f"Padded track {idx} for transcription: {padded_url}")
+
+        # Mixdown PADDED tracks (already aligned with timeline) into transcript.audio_mp3_filename
+        try:
+            # Ensure data directory exists
+            transcript.data_path.mkdir(parents=True, exist_ok=True)
+
+            mp3_writer = AudioFileWriterProcessor(
+                path=str(transcript.audio_mp3_filename),
+                on_duration=self.on_duration,
+            )
+            # Use PADDED tracks with NO additional offsets (already aligned by padding)
+            await self.mixdown_tracks(
+                padded_track_datas, mp3_writer, offsets_seconds=None
+            )
+            await mp3_writer.flush()
+
+            # Upload the mixed audio to S3 for web playback
+            if transcript.audio_mp3_filename.exists():
+                mp3_data = transcript.audio_mp3_filename.read_bytes()
+                storage_path = f"{transcript.id}/audio.mp3"
+                await storage.put_file(storage_path, mp3_data)
+                mp3_url = await storage.get_file_url(storage_path)
+
+                # Update transcript to indicate audio is in storage
+                await transcripts_controller.update(
+                    transcript, {"audio_location": "storage"}
+                )
+
+                self.logger.info(
+                    f"Uploaded mixed audio to storage",
+                    storage_path=storage_path,
+                    size=len(mp3_data),
+                    url=mp3_url,
+                )
+            else:
+                self.logger.warning("Mixdown file does not exist after processing")
+        except Exception as e:
+            self.logger.error("Mixdown failed", error=str(e), exc_info=True)
+
+        # Generate waveform from the mixed audio file
+        if transcript.audio_mp3_filename.exists():
+            try:
+                self.logger.info("Generating waveform from mixed audio")
+                waveform_processor = AudioWaveformProcessor(
+                    audio_path=transcript.audio_mp3_filename,
+                    waveform_path=transcript.audio_waveform_filename,
+                    on_waveform=self.on_waveform,
+                )
+                waveform_processor.set_pipeline(self.empty_pipeline)
+                await waveform_processor.flush()
+                self.logger.info("Waveform generated successfully")
+            except Exception as e:
+                self.logger.error(
+                    "Waveform generation failed", error=str(e), exc_info=True
+                )
+
+        # Transcribe PADDED tracks - timestamps will be automatically correct!
+        speaker_transcripts: list[TranscriptType] = []
+        for idx, padded_url in enumerate(padded_track_urls):
+            if not padded_url:
+                continue
+
+            try:
+                # Transcribe the PADDED track
+                t = await self.transcribe_file(padded_url, transcript.source_language)
+            except Exception as e:
+                self.logger.error(
+                    "Transcription via default backend failed, trying local whisper",
+                    track_idx=idx,
+                    url=padded_url,
+                    error=str(e),
+                )
+                try:
+                    fallback = FileTranscriptAutoProcessor(name="whisper")
+                    result = None
+
+                    async def capture_result(r):
+                        nonlocal result
+                        result = r
+
+                    fallback.on(capture_result)
+                    await fallback.push(
+                        FileTranscriptInput(
+                            audio_url=padded_url, language=transcript.source_language
+                        )
+                    )
+                    await fallback.flush()
+                    if not result:
+                        raise Exception("No transcript captured in fallback")
+                    t = result
+                except Exception as e2:
+                    self.logger.error(
+                        "Skipping track - transcription failed after fallback",
+                        track_idx=idx,
+                        url=padded_url,
+                        error=str(e2),
+                    )
+                    continue
+
+            if not t.words:
+                continue
+
+            # NO OFFSET ADJUSTMENT NEEDED!
+            # Timestamps are already correct because we transcribed padded tracks
+            # Just set speaker ID
+            for w in t.words:
+                w.speaker = idx
+
+            speaker_transcripts.append(t)
+            self.logger.info(
+                f"Track {idx} transcribed successfully with {len(t.words)} words",
+                track_idx=idx,
+            )
+
+        if not speaker_transcripts:
+            raise Exception("No valid track transcriptions")
+
+        # Merge all words and sort by timestamp
+        merged_words = []
+        for t in speaker_transcripts:
+            merged_words.extend(t.words)
+        merged_words.sort(
+            key=lambda w: w.start if hasattr(w, "start") and w.start is not None else 0
+        )
+
+        merged_transcript = TranscriptType(words=merged_words, translation=None)
+
+        # Emit TRANSCRIPT event through the shared handler (persists and broadcasts)
+        await self.on_transcript(merged_transcript)
+
+        topics = await self.detect_topics(merged_transcript, transcript.target_language)
+        await asyncio.gather(
+            self.generate_title(topics),
+            self.generate_summaries(topics),
+            return_exceptions=False,
+        )
+
+        await self.set_status(transcript.id, "ended")
+
+    async def transcribe_file(self, audio_url: str, language: str) -> TranscriptType:
+        processor = FileTranscriptAutoProcessor()
+        input_data = FileTranscriptInput(audio_url=audio_url, language=language)
+
+        result: TranscriptType | None = None
+
+        async def capture_result(transcript):
+            nonlocal result
+            result = transcript
+
+        processor.on(capture_result)
+        await processor.push(input_data)
+        await processor.flush()
+
+        if not result:
+            raise ValueError("No transcript captured")
+
+        return result
+
+    async def detect_topics(
+        self, transcript: TranscriptType, target_language: str
+    ) -> list[TitleSummary]:
+        chunk_size = 300
+        topics: list[TitleSummary] = []
+
+        async def on_topic(topic: TitleSummary):
+            topics.append(topic)
+            return await self.on_topic(topic)
+
+        topic_detector = TranscriptTopicDetectorProcessor(callback=on_topic)
+        topic_detector.set_pipeline(self.empty_pipeline)
+
+        for i in range(0, len(transcript.words), chunk_size):
+            chunk_words = transcript.words[i : i + chunk_size]
+            if not chunk_words:
+                continue
+
+            chunk_transcript = TranscriptType(
+                words=chunk_words, translation=transcript.translation
+            )
+            await topic_detector.push(chunk_transcript)
+
+        await topic_detector.flush()
+        return topics
+
+    async def generate_title(self, topics: list[TitleSummary]):
+        if not topics:
+            self.logger.warning("No topics for title generation")
+            return
+
+        processor = TranscriptFinalTitleProcessor(callback=self.on_title)
+        processor.set_pipeline(self.empty_pipeline)
+
+        for topic in topics:
+            await processor.push(topic)
+
+        await processor.flush()
+
+    async def generate_summaries(self, topics: list[TitleSummary]):
+        if not topics:
+            self.logger.warning("No topics for summary generation")
+            return
+
+        transcript = await self.get_transcript()
+        processor = TranscriptFinalSummaryProcessor(
+            transcript=transcript,
+            callback=self.on_long_summary,
+            on_short_summary=self.on_short_summary,
+        )
+        processor.set_pipeline(self.empty_pipeline)
+
+        for topic in topics:
+            await processor.push(topic)
+
+        await processor.flush()
+
+
+@shared_task
+@asynctask
+async def task_pipeline_multitrack_process(
+    *, transcript_id: str, bucket_name: str, track_keys: list[str]
+):
+    pipeline = PipelineMainMultitrack(transcript_id=transcript_id)
+    try:
+        await pipeline.set_status(transcript_id, "processing")
+        await pipeline.process(bucket_name, track_keys)
+    except Exception:
+        await pipeline.set_status(transcript_id, "error")
+        raise
+
+    post_chain = chain(
+        task_cleanup_consent.si(transcript_id=transcript_id),
+        task_pipeline_post_to_zulip.si(transcript_id=transcript_id),
+        task_send_webhook_if_needed.si(transcript_id=transcript_id),
+    )
+    post_chain.delay()

server/reflector/pipelines/main_multitrack_pipeline_fixed.py

@@ -0,0 +1,629 @@
+import asyncio
+import io
+from fractions import Fraction
+
+import av
+import boto3
+import structlog
+from av.audio.resampler import AudioResampler
+from celery import chain, shared_task
+
+from reflector.asynctask import asynctask
+from reflector.db.transcripts import (
+    TranscriptStatus,
+    TranscriptText,
+    transcripts_controller,
+)
+from reflector.logger import logger
+from reflector.pipelines.main_file_pipeline import task_send_webhook_if_needed
+from reflector.pipelines.main_live_pipeline import (
+    PipelineMainBase,
+    task_cleanup_consent,
+    task_pipeline_post_to_zulip,
+)
+from reflector.processors import (
+    AudioFileWriterProcessor,
+    TranscriptFinalSummaryProcessor,
+    TranscriptFinalTitleProcessor,
+    TranscriptTopicDetectorProcessor,
+)
+from reflector.processors.file_transcript import FileTranscriptInput
+from reflector.processors.file_transcript_auto import FileTranscriptAutoProcessor
+from reflector.processors.types import TitleSummary
+from reflector.processors.types import (
+    Transcript as TranscriptType,
+)
+from reflector.settings import settings
+from reflector.storage import get_transcripts_storage
+
+
+class EmptyPipeline:
+    def __init__(self, logger: structlog.BoundLogger):
+        self.logger = logger
+
+    def get_pref(self, k, d=None):
+        return d
+
+    async def emit(self, event):
+        pass
+
+
+class PipelineMainMultitrack(PipelineMainBase):
+    """Process multiple participant tracks for a transcript without mixing audio."""
+
+    def __init__(self, transcript_id: str):
+        super().__init__(transcript_id=transcript_id)
+        self.logger = logger.bind(transcript_id=self.transcript_id)
+        self.empty_pipeline = EmptyPipeline(logger=self.logger)
+
+    async def pad_track_for_transcription(
+        self,
+        track_data: bytes,
+        track_idx: int,
+        storage,
+    ) -> tuple[bytes, str]:
+        """
+        Pad a single track with silence based on stream metadata start_time.
+        This ensures Whisper timestamps will be relative to recording start.
+
+        Returns: (padded_data, storage_url)
+        """
+        if not track_data:
+            return b"", ""
+
+        transcript = await self.get_transcript()
+
+        # Get stream metadata start_time using PyAV
+        container = av.open(io.BytesIO(track_data))
+        try:
+            audio_stream = container.streams.audio[0]
+
+            # Extract start_time from stream metadata
+            if (
+                audio_stream.start_time is not None
+                and audio_stream.time_base is not None
+            ):
+                start_time_seconds = float(
+                    audio_stream.start_time * audio_stream.time_base
+                )
+            else:
+                start_time_seconds = 0.0
+
+            sample_rate = audio_stream.sample_rate
+            codec_name = audio_stream.codec.name
+        finally:
+            container.close()
+
+        self.logger.info(
+            f"Track {track_idx} stream metadata: start_time={start_time_seconds:.3f}s, sample_rate={sample_rate}",
+            track_idx=track_idx,
+        )
+
+        # If no padding needed, use original
+        if start_time_seconds <= 0:
+            storage_path = (
+                f"file_pipeline/{transcript.id}/tracks/original_track_{track_idx}.webm"
+            )
+            await storage.put_file(storage_path, track_data)
+            url = await storage.get_file_url(storage_path)
+            return track_data, url
+
+        # Create PyAV filter graph for padding
+        graph = av.filter.Graph()
+
+        # Input buffer
+        in_args = (
+            f"time_base=1/{sample_rate}:"
+            f"sample_rate={sample_rate}:"
+            f"sample_fmt=s16:"
+            f"channel_layout=stereo"
+        )
+        input_buffer = graph.add("abuffer", args=in_args, name="in")
+
+        # Add delay filter for padding
+        delay_ms = int(start_time_seconds * 1000)
+        delay_filter = graph.add(
+            "adelay", args=f"delays={delay_ms}|{delay_ms}:all=1", name="delay"
+        )
+
+        # Output sink
+        sink = graph.add("abuffersink", name="out")
+
+        # Link filters
+        input_buffer.link_to(delay_filter)
+        delay_filter.link_to(sink)
+
+        graph.configure()
+
+        # Process audio through filter
+        output_bytes = io.BytesIO()
+        output_container = av.open(output_bytes, "w", format="webm")
+        output_stream = output_container.add_stream("libopus", rate=sample_rate)
+        output_stream.channels = 2
+
+        # Reopen input for processing
+        input_container = av.open(io.BytesIO(track_data))
+        resampler = AudioResampler(format="s16", layout="stereo", rate=sample_rate)
+
+        try:
+            # Process frames
+            for frame in input_container.decode(audio=0):
+                # Resample to match filter requirements
+                resampled_frames = resampler.resample(frame)
+                for resampled_frame in resampled_frames:
+                    resampled_frame.pts = frame.pts
+                    resampled_frame.time_base = Fraction(1, sample_rate)
+                    input_buffer.push(resampled_frame)
+
+                    # Pull from filter and encode
+                    while True:
+                        try:
+                            out_frame = sink.pull()
+                            out_frame.pts = out_frame.pts if out_frame.pts else 0
+                            out_frame.time_base = Fraction(1, sample_rate)
+                            for packet in output_stream.encode(out_frame):
+                                output_container.mux(packet)
+                        except av.BlockingIOError:
+                            break
+
+            # Flush
+            input_buffer.push(None)
+            while True:
+                try:
+                    out_frame = sink.pull()
+                    for packet in output_stream.encode(out_frame):
+                        output_container.mux(packet)
+                except (av.BlockingIOError, av.EOFError):
+                    break
+
+            # Flush encoder
+            for packet in output_stream.encode(None):
+                output_container.mux(packet)
+
+        finally:
+            input_container.close()
+            output_container.close()
+
+        padded_data = output_bytes.getvalue()
+
+        # Store padded track
+        storage_path = (
+            f"file_pipeline/{transcript.id}/tracks/padded_track_{track_idx}.webm"
+        )
+        await storage.put_file(storage_path, padded_data)
+        padded_url = await storage.get_file_url(storage_path)
+
+        self.logger.info(
+            f"Padded track {track_idx} with {start_time_seconds:.3f}s offset, stored at {storage_path}",
+            track_idx=track_idx,
+            delay_ms=delay_ms,
+            padded_url=padded_url,
+        )
+
+        return padded_data, padded_url
+
+    async def mixdown_tracks(
+        self,
+        track_datas: list[bytes],
+        writer: AudioFileWriterProcessor,
+        offsets_seconds: list[float] | None = None,
+    ) -> None:
+        """
+        Minimal multi-track mixdown using a PyAV filter graph (amix), no resampling.
+        """
+
+        # Discover target sample rate from first decodable frame
+        target_sample_rate: int | None = None
+        for data in track_datas:
+            if not data:
+                continue
+            try:
+                container = av.open(io.BytesIO(data))
+                try:
+                    for frame in container.decode(audio=0):
+                        target_sample_rate = frame.sample_rate
+                        break
+                finally:
+                    container.close()
+            except Exception:
+                continue
+            if target_sample_rate:
+                break
+
+        if not target_sample_rate:
+            self.logger.warning("Mixdown skipped - no decodable audio frames found")
+            return
+
+        # Build PyAV filter graph:
+        # N abuffer (s32/stereo)
+        #   -> optional adelay per input (for alignment)
+        #   -> amix (s32)
+        #   -> aformat(s16)
+        #   -> sink
+        graph = av.filter.Graph()
+        inputs = []
+        valid_track_datas = [d for d in track_datas if d]
+        # Align offsets list with the filtered inputs (skip empties)
+        input_offsets_seconds = None
+        if offsets_seconds is not None:
+            input_offsets_seconds = [
+                offsets_seconds[i] for i, d in enumerate(track_datas) if d
+            ]
+        for idx, data in enumerate(valid_track_datas):
+            args = (
+                f"time_base=1/{target_sample_rate}:"
+                f"sample_rate={target_sample_rate}:"
+                f"sample_fmt=s32:"
+                f"channel_layout=stereo"
+            )
+            in_ctx = graph.add("abuffer", args=args, name=f"in{idx}")
+            inputs.append(in_ctx)
+
+        if not inputs:
+            self.logger.warning("Mixdown skipped - no valid inputs for graph")
+            return
+
+        mixer = graph.add("amix", args=f"inputs={len(inputs)}:normalize=0", name="mix")
+
+        fmt = graph.add(
+            "aformat",
+            args=(
+                f"sample_fmts=s32:channel_layouts=stereo:sample_rates={target_sample_rate}"
+            ),
+            name="fmt",
+        )
+
+        sink = graph.add("abuffersink", name="out")
+
+        # Optional per-input delay before mixing
+        delays_ms: list[int] = []
+        if input_offsets_seconds is not None:
+            base = min(input_offsets_seconds) if input_offsets_seconds else 0.0
+            delays_ms = [
+                max(0, int(round((o - base) * 1000))) for o in input_offsets_seconds
+            ]
+        else:
+            delays_ms = [0 for _ in inputs]
+
+        for idx, in_ctx in enumerate(inputs):
+            delay_ms = delays_ms[idx] if idx < len(delays_ms) else 0
+            if delay_ms > 0:
+                # adelay requires one value per channel; use same for stereo
+                adelay = graph.add(
+                    "adelay",
+                    args=f"delays={delay_ms}|{delay_ms}:all=1",
+                    name=f"delay{idx}",
+                )
+                in_ctx.link_to(adelay)
+                adelay.link_to(mixer, 0, idx)
+            else:
+                in_ctx.link_to(mixer, 0, idx)
+        mixer.link_to(fmt)
+        fmt.link_to(sink)
+        graph.configure()
+
+        # Open containers for decoding
+        containers = []
+        for i, d in enumerate(valid_track_datas):
+            try:
+                c = av.open(io.BytesIO(d))
+                containers.append(c)
+            except Exception as e:
+                self.logger.warning(
+                    "Mixdown: failed to open container", input=i, error=str(e)
+                )
+                containers.append(None)
+        # Filter out Nones for decoders
+        containers = [c for c in containers if c is not None]
+        decoders = [c.decode(audio=0) for c in containers]
+        active = [True] * len(decoders)
+        # Per-input resamplers to enforce s32/stereo at the same rate (no resample of rate)
+        resamplers = [
+            AudioResampler(format="s32", layout="stereo", rate=target_sample_rate)
+            for _ in decoders
+        ]
+
+        try:
+            # Round-robin feed frames into graph, pull mixed frames as they become available
+            while any(active):
+                for i, (dec, is_active) in enumerate(zip(decoders, active)):
+                    if not is_active:
+                        continue
+                    try:
+                        frame = next(dec)
+                    except StopIteration:
+                        active[i] = False
+                        continue
+
+                    # Enforce same sample rate; convert format/layout to s16/stereo (no resample)
+                    if frame.sample_rate != target_sample_rate:
+                        # Skip frames with differing rate
+                        continue
+                    out_frames = resamplers[i].resample(frame) or []
+                    for rf in out_frames:
+                        rf.sample_rate = target_sample_rate
+                        rf.time_base = Fraction(1, target_sample_rate)
+                        inputs[i].push(rf)
+
+                    # Drain available mixed frames
+                    while True:
+                        try:
+                            mixed = sink.pull()
+                        except Exception:
+                            break
+                        mixed.sample_rate = target_sample_rate
+                        mixed.time_base = Fraction(1, target_sample_rate)
+                        await writer.push(mixed)
+
+            # Signal EOF to inputs and drain remaining
+            for in_ctx in inputs:
+                in_ctx.push(None)
+            while True:
+                try:
+                    mixed = sink.pull()
+                except Exception:
+                    break
+                mixed.sample_rate = target_sample_rate
+                mixed.time_base = Fraction(1, target_sample_rate)
+                await writer.push(mixed)
+        finally:
+            for c in containers:
+                c.close()
+
+    async def set_status(self, transcript_id: str, status: TranscriptStatus):
+        async with self.lock_transaction():
+            return await transcripts_controller.set_status(transcript_id, status)
+
+    async def process(self, bucket_name: str, track_keys: list[str]):
+        transcript = await self.get_transcript()
+
+        s3 = boto3.client(
+            "s3",
+            region_name=settings.RECORDING_STORAGE_AWS_REGION,
+            aws_access_key_id=settings.RECORDING_STORAGE_AWS_ACCESS_KEY_ID,
+            aws_secret_access_key=settings.RECORDING_STORAGE_AWS_SECRET_ACCESS_KEY,
+        )
+
+        storage = get_transcripts_storage()
+
+        # Pre-download bytes for all tracks for mixing and transcription
+        track_datas: list[bytes] = []
+        for key in track_keys:
+            try:
+                obj = s3.get_object(Bucket=bucket_name, Key=key)
+                track_datas.append(obj["Body"].read())
+            except Exception as e:
+                self.logger.warning(
+                    "Skipping track - cannot read S3 object", key=key, error=str(e)
+                )
+                track_datas.append(b"")
+
+        # REMOVED: Filename offset extraction - not needed anymore!
+        # We use stream metadata start_time for padding instead
+
+        # Get stream metadata start_times for mixing (still useful for mixdown)
+        stream_start_times: list[float] = []
+        for data in track_datas:
+            if not data:
+                stream_start_times.append(0.0)
+                continue
+
+            container = av.open(io.BytesIO(data))
+            try:
+                audio_stream = container.streams.audio[0]
+                if (
+                    audio_stream.start_time is not None
+                    and audio_stream.time_base is not None
+                ):
+                    start_time = float(audio_stream.start_time * audio_stream.time_base)
+                else:
+                    start_time = 0.0
+                stream_start_times.append(start_time)
+            finally:
+                container.close()
+
+        # Mixdown all available tracks into transcript.audio_mp3_filename, using stream metadata offsets
+        try:
+            mp3_writer = AudioFileWriterProcessor(
+                path=str(transcript.audio_mp3_filename)
+            )
+            await self.mixdown_tracks(track_datas, mp3_writer, stream_start_times)
+            await mp3_writer.flush()
+        except Exception as e:
+            self.logger.error("Mixdown failed", error=str(e))
+
+        # PAD TRACKS BEFORE TRANSCRIPTION - THIS IS THE KEY FIX!
+        padded_track_urls: list[str] = []
+        for idx, data in enumerate(track_datas):
+            if not data:
+                padded_track_urls.append("")
+                continue
+
+            _, padded_url = await self.pad_track_for_transcription(data, idx, storage)
+            padded_track_urls.append(padded_url)
+            self.logger.info(f"Padded track {idx} for transcription: {padded_url}")
+
+        # Transcribe PADDED tracks - timestamps will be automatically correct!
+        speaker_transcripts: list[TranscriptType] = []
+        for idx, padded_url in enumerate(padded_track_urls):
+            if not padded_url:
+                continue
+
+            try:
+                # Transcribe the PADDED track
+                t = await self.transcribe_file(padded_url, transcript.source_language)
+            except Exception as e:
+                self.logger.error(
+                    "Transcription via default backend failed, trying local whisper",
+                    track_idx=idx,
+                    url=padded_url,
+                    error=str(e),
+                )
+                try:
+                    fallback = FileTranscriptAutoProcessor(name="whisper")
+                    result = None
+
+                    async def capture_result(r):
+                        nonlocal result
+                        result = r
+
+                    fallback.on(capture_result)
+                    await fallback.push(
+                        FileTranscriptInput(
+                            audio_url=padded_url, language=transcript.source_language
+                        )
+                    )
+                    await fallback.flush()
+                    if not result:
+                        raise Exception("No transcript captured in fallback")
+                    t = result
+                except Exception as e2:
+                    self.logger.error(
+                        "Skipping track - transcription failed after fallback",
+                        track_idx=idx,
+                        url=padded_url,
+                        error=str(e2),
+                    )
+                    continue
+
+            if not t.words:
+                continue
+
+            # NO OFFSET ADJUSTMENT NEEDED!
+            # Timestamps are already correct because we transcribed padded tracks
+            # Just set speaker ID
+            for w in t.words:
+                w.speaker = idx
+
+            speaker_transcripts.append(t)
+            self.logger.info(
+                f"Track {idx} transcribed successfully with {len(t.words)} words",
+                track_idx=idx,
+            )
+
+        if not speaker_transcripts:
+            raise Exception("No valid track transcriptions")
+
+        # Merge all words and sort by timestamp
+        merged_words = []
+        for t in speaker_transcripts:
+            merged_words.extend(t.words)
+        merged_words.sort(
+            key=lambda w: w.start if hasattr(w, "start") and w.start is not None else 0
+        )
+
+        merged_transcript = TranscriptType(words=merged_words, translation=None)
+
+        await transcripts_controller.append_event(
+            transcript,
+            event="TRANSCRIPT",
+            data=TranscriptText(
+                text=merged_transcript.text, translation=merged_transcript.translation
+            ),
+        )
+
+        topics = await self.detect_topics(merged_transcript, transcript.target_language)
+        await asyncio.gather(
+            self.generate_title(topics),
+            self.generate_summaries(topics),
+            return_exceptions=False,
+        )
+
+        await self.set_status(transcript.id, "ended")
+
+    async def transcribe_file(self, audio_url: str, language: str) -> TranscriptType:
+        processor = FileTranscriptAutoProcessor()
+        input_data = FileTranscriptInput(audio_url=audio_url, language=language)
+
+        result: TranscriptType | None = None
+
+        async def capture_result(transcript):
+            nonlocal result
+            result = transcript
+
+        processor.on(capture_result)
+        await processor.push(input_data)
+        await processor.flush()
+
+        if not result:
+            raise ValueError("No transcript captured")
+
+        return result
+
+    async def detect_topics(
+        self, transcript: TranscriptType, target_language: str
+    ) -> list[TitleSummary]:
+        chunk_size = 300
+        topics: list[TitleSummary] = []
+
+        async def on_topic(topic: TitleSummary):
+            topics.append(topic)
+            return await self.on_topic(topic)
+
+        topic_detector = TranscriptTopicDetectorProcessor(callback=on_topic)
+        topic_detector.set_pipeline(self.empty_pipeline)
+
+        for i in range(0, len(transcript.words), chunk_size):
+            chunk_words = transcript.words[i : i + chunk_size]
+            if not chunk_words:
+                continue
+
+            chunk_transcript = TranscriptType(
+                words=chunk_words, translation=transcript.translation
+            )
+            await topic_detector.push(chunk_transcript)
+
+        await topic_detector.flush()
+        return topics
+
+    async def generate_title(self, topics: list[TitleSummary]):
+        if not topics:
+            self.logger.warning("No topics for title generation")
+            return
+
+        processor = TranscriptFinalTitleProcessor(callback=self.on_title)
+        processor.set_pipeline(self.empty_pipeline)
+
+        for topic in topics:
+            await processor.push(topic)
+
+        await processor.flush()
+
+    async def generate_summaries(self, topics: list[TitleSummary]):
+        if not topics:
+            self.logger.warning("No topics for summary generation")
+            return
+
+        transcript = await self.get_transcript()
+        processor = TranscriptFinalSummaryProcessor(
+            transcript=transcript,
+            callback=self.on_long_summary,
+            on_short_summary=self.on_short_summary,
+        )
+        processor.set_pipeline(self.empty_pipeline)
+
+        for topic in topics:
+            await processor.push(topic)
+
+        await processor.flush()
+
+
+@shared_task
+@asynctask
+async def task_pipeline_multitrack_process(
+    *, transcript_id: str, bucket_name: str, track_keys: list[str]
+):
+    pipeline = PipelineMainMultitrack(transcript_id=transcript_id)
+    try:
+        await pipeline.set_status(transcript_id, "processing")
+        await pipeline.process(bucket_name, track_keys)
+    except Exception:
+        await pipeline.set_status(transcript_id, "error")
+        raise
+
+    post_chain = chain(
+        task_cleanup_consent.si(transcript_id=transcript_id),
+        task_pipeline_post_to_zulip.si(transcript_id=transcript_id),
+        task_send_webhook_if_needed.si(transcript_id=transcript_id),
+    )
+    post_chain.delay()

server/reflector/platform_types.py

@@ -0,0 +1,9 @@
+"""Platform type definitions.
+
+This module exists solely to define the Platform literal type without any imports,
+preventing circular import issues when used across the codebase.
+"""
+
+from typing import Literal
+
+Platform = Literal["whereby", "daily"]

server/reflector/processors/__init__.py

@@ -1,5 +1,7 @@
 from .audio_chunker import AudioChunkerProcessor  # noqa: F401
+from .audio_chunker_auto import AudioChunkerAutoProcessor  # noqa: F401
 from .audio_diarization_auto import AudioDiarizationAutoProcessor  # noqa: F401
+from .audio_downscale import AudioDownscaleProcessor  # noqa: F401
 from .audio_file_writer import AudioFileWriterProcessor  # noqa: F401
 from .audio_merge import AudioMergeProcessor  # noqa: F401
 from .audio_transcript import AudioTranscriptProcessor  # noqa: F401

server/reflector/processors/audio_chunker.py

@@ -1,340 +1,78 @@
 from typing import Optional
 
 import av
-import numpy as np
-import torch
-from silero_vad import VADIterator, load_silero_vad
+from prometheus_client import Counter, Histogram
 
 from reflector.processors.base import Processor
 
 
 class AudioChunkerProcessor(Processor):
     """
-    Assemble audio frames into chunks with VAD-based speech detection
+    Base class for assembling audio frames into chunks
     """
 
     INPUT_TYPE = av.AudioFrame
     OUTPUT_TYPE = list[av.AudioFrame]
 
-    def __init__(
-        self,
-        block_frames=256,
-        max_frames=1024,
-        vad_threshold=0.5,
-        use_onnx=False,
-        min_frames=2,
-    ):
-        super().__init__()
+    m_chunk = Histogram(
+        "audio_chunker",
+        "Time spent in AudioChunker.chunk",
+        ["backend"],
+    )
+    m_chunk_call = Counter(
+        "audio_chunker_call",
+        "Number of calls to AudioChunker.chunk",
+        ["backend"],
+    )
+    m_chunk_success = Counter(
+        "audio_chunker_success",
+        "Number of successful calls to AudioChunker.chunk",
+        ["backend"],
+    )
+    m_chunk_failure = Counter(
+        "audio_chunker_failure",
+        "Number of failed calls to AudioChunker.chunk",
+        ["backend"],
+    )
+
+    def __init__(self, *args, **kwargs):
+        name = self.__class__.__name__
+        self.m_chunk = self.m_chunk.labels(name)
+        self.m_chunk_call = self.m_chunk_call.labels(name)
+        self.m_chunk_success = self.m_chunk_success.labels(name)
+        self.m_chunk_failure = self.m_chunk_failure.labels(name)
+        super().__init__(*args, **kwargs)
         self.frames: list[av.AudioFrame] = []
-        self.block_frames = block_frames
-        self.max_frames = max_frames
-        self.vad_threshold = vad_threshold
-        self.min_frames = min_frames
-
-        # Initialize Silero VAD
-        self._init_vad(use_onnx)
-
-    def _init_vad(self, use_onnx=False):
-        """Initialize Silero VAD model"""
-        try:
-            torch.set_num_threads(1)
-            self.vad_model = load_silero_vad(onnx=use_onnx)
-            self.vad_iterator = VADIterator(self.vad_model, sampling_rate=16000)
-            self.logger.info("Silero VAD initialized successfully")
-
-        except Exception as e:
-            self.logger.error(f"Failed to initialize Silero VAD: {e}")
-            self.vad_model = None
-            self.vad_iterator = None
 
     async def _push(self, data: av.AudioFrame):
-        self.frames.append(data)
-        # print("timestamp", data.pts * data.time_base * 1000)
-
-        # Check for speech segments every 32 frames (~1 second)
-        if len(self.frames) >= 32 and len(self.frames) % 32 == 0:
-            await self._process_block()
-
-        # Safety fallback - emit if we hit max frames
-        elif len(self.frames) >= self.max_frames:
-            self.logger.warning(
-                f"AudioChunkerProcessor: Reached max frames ({self.max_frames}), "
-                f"emitting first {self.max_frames // 2} frames"
-            )
-            frames_to_emit = self.frames[: self.max_frames // 2]
-            self.frames = self.frames[self.max_frames // 2 :]
-            if len(frames_to_emit) >= self.min_frames:
-                await self.emit(frames_to_emit)
-            else:
-                self.logger.debug(
-                    f"Ignoring fallback segment with {len(frames_to_emit)} frames "
-                    f"(< {self.min_frames} minimum)"
+        """Process incoming audio frame"""
+        # Validate audio format on first frame
+        if len(self.frames) == 0:
+            if data.sample_rate != 16000 or len(data.layout.channels) != 1:
+                raise ValueError(
+                    f"AudioChunkerProcessor expects 16kHz mono audio, got {data.sample_rate}Hz "
+                    f"with {len(data.layout.channels)} channel(s). "
+                    f"Use AudioDownscaleProcessor before this processor."
                 )
 
-    async def _process_block(self):
-        # Need at least 32 frames for VAD detection (~1 second)
-        if len(self.frames) < 32 or self.vad_iterator is None:
-            return
-
-        # Processing block with current buffer size
-        # print(f"Processing block: {len(self.frames)} frames in buffer")
-
         try:
-            # Convert frames to numpy array for VAD
-            audio_array = self._frames_to_numpy(self.frames)
+            self.m_chunk_call.inc()
+            with self.m_chunk.time():
+                result = await self._chunk(data)
+            self.m_chunk_success.inc()
+            if result:
+                await self.emit(result)
+        except Exception:
+            self.m_chunk_failure.inc()
+            raise
 
-            if audio_array is None:
-                # Fallback: emit all frames if conversion failed
-                frames_to_emit = self.frames[:]
-                self.frames = []
-                if len(frames_to_emit) >= self.min_frames:
-                    await self.emit(frames_to_emit)
-                else:
-                    self.logger.debug(
-                        f"Ignoring conversion-failed segment with {len(frames_to_emit)} frames "
-                        f"(< {self.min_frames} minimum)"
-                    )
-                return
-
-            # Find complete speech segments in the buffer
-            speech_end_frame = self._find_speech_segment_end(audio_array)
-
-            if speech_end_frame is None or speech_end_frame <= 0:
-                # No speech found but buffer is getting large
-                if len(self.frames) > 512:
-                    # Check if it's all silence and can be discarded
-                    # No speech segment found, buffer at {len(self.frames)} frames
-
-                    # Could emit silence or discard old frames here
-                    # For now, keep first 256 frames and discard older silence
-                    if len(self.frames) > 768:
-                        self.logger.debug(
-                            f"Discarding {len(self.frames) - 256} old frames (likely silence)"
-                        )
-                        self.frames = self.frames[-256:]
-                return
-
-            # Calculate segment timing information
-            frames_to_emit = self.frames[:speech_end_frame]
-
-            # Get timing from av.AudioFrame
-            if frames_to_emit:
-                first_frame = frames_to_emit[0]
-                last_frame = frames_to_emit[-1]
-                sample_rate = first_frame.sample_rate
-
-                # Calculate duration
-                total_samples = sum(f.samples for f in frames_to_emit)
-                duration_seconds = total_samples / sample_rate if sample_rate > 0 else 0
-
-                # Get timestamps if available
-                start_time = (
-                    first_frame.pts * first_frame.time_base if first_frame.pts else 0
-                )
-                end_time = (
-                    last_frame.pts * last_frame.time_base if last_frame.pts else 0
-                )
-
-                # Convert to HH:MM:SS format for logging
-                def format_time(seconds):
-                    if not seconds:
-                        return "00:00:00"
-                    total_seconds = int(float(seconds))
-                    hours = total_seconds // 3600
-                    minutes = (total_seconds % 3600) // 60
-                    secs = total_seconds % 60
-                    return f"{hours:02d}:{minutes:02d}:{secs:02d}"
-
-                start_formatted = format_time(start_time)
-                end_formatted = format_time(end_time)
-
-                # Keep remaining frames for next processing
-                remaining_after = len(self.frames) - speech_end_frame
-
-                # Single structured log line
-                self.logger.info(
-                    "Speech segment found",
-                    start=start_formatted,
-                    end=end_formatted,
-                    frames=speech_end_frame,
-                    duration=round(duration_seconds, 2),
-                    buffer_before=len(self.frames),
-                    remaining=remaining_after,
-                )
-
-            # Keep remaining frames for next processing
-            self.frames = self.frames[speech_end_frame:]
-
-            # Filter out segments with too few frames
-            if len(frames_to_emit) >= self.min_frames:
-                await self.emit(frames_to_emit)
-            else:
-                self.logger.debug(
-                    f"Ignoring segment with {len(frames_to_emit)} frames "
-                    f"(< {self.min_frames} minimum)"
-                )
-
-        except Exception as e:
-            self.logger.error(f"Error in VAD processing: {e}")
-            # Fallback to simple chunking
-            if len(self.frames) >= self.block_frames:
-                frames_to_emit = self.frames[: self.block_frames]
-                self.frames = self.frames[self.block_frames :]
-                if len(frames_to_emit) >= self.min_frames:
-                    await self.emit(frames_to_emit)
-                else:
-                    self.logger.debug(
-                        f"Ignoring exception-fallback segment with {len(frames_to_emit)} frames "
-                        f"(< {self.min_frames} minimum)"
-                    )
-
-    def _frames_to_numpy(self, frames: list[av.AudioFrame]) -> Optional[np.ndarray]:
-        """Convert av.AudioFrame list to numpy array for VAD processing"""
-        if not frames:
-            return None
-
-        try:
-            first_frame = frames[0]
-            original_sample_rate = first_frame.sample_rate
-
-            audio_data = []
-            for frame in frames:
-                frame_array = frame.to_ndarray()
-
-                # Handle stereo -> mono conversion
-                if len(frame_array.shape) == 2 and frame_array.shape[0] > 1:
-                    frame_array = np.mean(frame_array, axis=0)
-                elif len(frame_array.shape) == 2:
-                    frame_array = frame_array.flatten()
-
-                audio_data.append(frame_array)
-
-            if not audio_data:
-                return None
-
-            combined_audio = np.concatenate(audio_data)
-
-            # Resample from 48kHz to 16kHz if needed
-            if original_sample_rate != 16000:
-                combined_audio = self._resample_audio(
-                    combined_audio, original_sample_rate, 16000
-                )
-
-            # Ensure float32 format
-            if combined_audio.dtype == np.int16:
-                # Normalize int16 audio to float32 in range [-1.0, 1.0]
-                combined_audio = combined_audio.astype(np.float32) / 32768.0
-            elif combined_audio.dtype != np.float32:
-                combined_audio = combined_audio.astype(np.float32)
-
-            return combined_audio
-
-        except Exception as e:
-            self.logger.error(f"Error converting frames to numpy: {e}")
-
-        return None
-
-    def _resample_audio(
-        self, audio: np.ndarray, from_sr: int, to_sr: int
-    ) -> np.ndarray:
-        """Simple linear resampling from from_sr to to_sr"""
-        if from_sr == to_sr:
-            return audio
-
-        try:
-            # Simple linear interpolation resampling
-            ratio = to_sr / from_sr
-            new_length = int(len(audio) * ratio)
-
-            # Create indices for interpolation
-            old_indices = np.linspace(0, len(audio) - 1, new_length)
-            resampled = np.interp(old_indices, np.arange(len(audio)), audio)
-
-            return resampled.astype(np.float32)
-
-        except Exception as e:
-            self.logger.error("Resampling error", exc_info=e)
-            # Fallback: simple decimation/repetition
-            if from_sr > to_sr:
-                # Downsample by taking every nth sample
-                step = from_sr // to_sr
-                return audio[::step]
-            else:
-                # Upsample by repeating samples
-                repeat = to_sr // from_sr
-                return np.repeat(audio, repeat)
-
-    def _find_speech_segment_end(self, audio_array: np.ndarray) -> Optional[int]:
-        """Find complete speech segments and return frame index at segment end"""
-        if self.vad_iterator is None or len(audio_array) == 0:
-            return None
-
-        try:
-            # Process audio in 512-sample windows for VAD
-            window_size = 512
-            min_silence_windows = 3  # Require 3 windows of silence after speech
-
-            # Track speech state
-            in_speech = False
-            speech_start = None
-            speech_end = None
-            silence_count = 0
-
-            for i in range(0, len(audio_array), window_size):
-                chunk = audio_array[i : i + window_size]
-                if len(chunk) < window_size:
-                    chunk = np.pad(chunk, (0, window_size - len(chunk)))
-
-                # Detect if this window has speech
-                speech_dict = self.vad_iterator(chunk, return_seconds=True)
-
-                # VADIterator returns dict with 'start' and 'end' when speech segments are detected
-                if speech_dict:
-                    if not in_speech:
-                        # Speech started
-                        speech_start = i
-                        in_speech = True
-                        # Debug: print(f"Speech START at sample {i}, VAD: {speech_dict}")
-                    silence_count = 0  # Reset silence counter
-                    continue
-
-                if not in_speech:
-                    continue
-
-                # We're in speech but found silence
-                silence_count += 1
-                if silence_count < min_silence_windows:
-                    continue
-
-                # Found end of speech segment
-                speech_end = i - (min_silence_windows - 1) * window_size
-                # Debug: print(f"Speech END at sample {speech_end}")
-
-                # Convert sample position to frame index
-                samples_per_frame = self.frames[0].samples if self.frames else 1024
-                # Account for resampling: we process at 16kHz but frames might be 48kHz
-                resample_ratio = 48000 / 16000  # 3x
-                actual_sample_pos = int(speech_end * resample_ratio)
-                frame_index = actual_sample_pos // samples_per_frame
-
-                # Ensure we don't exceed buffer
-                frame_index = min(frame_index, len(self.frames))
-                return frame_index
-
-            return None
-
-        except Exception as e:
-            self.logger.error(f"Error finding speech segment: {e}")
-            return None
+    async def _chunk(self, data: av.AudioFrame) -> Optional[list[av.AudioFrame]]:
+        """
+        Process audio frame and return chunk when ready.
+        Subclasses should implement their chunking logic here.
+        """
+        raise NotImplementedError
 
     async def _flush(self):
-        frames = self.frames[:]
-        self.frames = []
-        if frames:
-            if len(frames) >= self.min_frames:
-                await self.emit(frames)
-            else:
-                self.logger.debug(
-                    f"Ignoring flush segment with {len(frames)} frames "
-                    f"(< {self.min_frames} minimum)"
-                )
+        """Flush any remaining frames when processing ends"""
+        raise NotImplementedError

server/reflector/processors/audio_chunker_auto.py

@@ -0,0 +1,32 @@
+import importlib
+
+from reflector.processors.audio_chunker import AudioChunkerProcessor
+from reflector.settings import settings
+
+
+class AudioChunkerAutoProcessor(AudioChunkerProcessor):
+    _registry = {}
+
+    @classmethod
+    def register(cls, name, kclass):
+        cls._registry[name] = kclass
+
+    def __new__(cls, name: str | None = None, **kwargs):
+        if name is None:
+            name = settings.AUDIO_CHUNKER_BACKEND
+        if name not in cls._registry:
+            module_name = f"reflector.processors.audio_chunker_{name}"
+            importlib.import_module(module_name)
+
+        # gather specific configuration for the processor
+        # search `AUDIO_CHUNKER_BACKEND_XXX_YYY`, push to constructor as `backend_xxx_yyy`
+        config = {}
+        name_upper = name.upper()
+        settings_prefix = "AUDIO_CHUNKER_"
+        config_prefix = f"{settings_prefix}{name_upper}_"
+        for key, value in settings:
+            if key.startswith(config_prefix):
+                config_name = key[len(settings_prefix) :].lower()
+                config[config_name] = value
+
+        return cls._registry[name](**config | kwargs)

server/reflector/processors/audio_chunker_frames.py

@@ -0,0 +1,34 @@
+from typing import Optional
+
+import av
+
+from reflector.processors.audio_chunker import AudioChunkerProcessor
+from reflector.processors.audio_chunker_auto import AudioChunkerAutoProcessor
+
+
+class AudioChunkerFramesProcessor(AudioChunkerProcessor):
+    """
+    Simple frame-based audio chunker that emits chunks after a fixed number of frames
+    """
+
+    def __init__(self, max_frames=256, **kwargs):
+        super().__init__(**kwargs)
+        self.max_frames = max_frames
+
+    async def _chunk(self, data: av.AudioFrame) -> Optional[list[av.AudioFrame]]:
+        self.frames.append(data)
+        if len(self.frames) >= self.max_frames:
+            frames_to_emit = self.frames[:]
+            self.frames = []
+            return frames_to_emit
+
+        return None
+
+    async def _flush(self):
+        frames = self.frames[:]
+        self.frames = []
+        if frames:
+            await self.emit(frames)
+
+
+AudioChunkerAutoProcessor.register("frames", AudioChunkerFramesProcessor)

server/reflector/processors/audio_chunker_silero.py

@@ -0,0 +1,298 @@
+from typing import Optional
+
+import av
+import numpy as np
+import torch
+from silero_vad import VADIterator, load_silero_vad
+
+from reflector.processors.audio_chunker import AudioChunkerProcessor
+from reflector.processors.audio_chunker_auto import AudioChunkerAutoProcessor
+
+
+class AudioChunkerSileroProcessor(AudioChunkerProcessor):
+    """
+    Assemble audio frames into chunks with VAD-based speech detection using Silero VAD
+    """
+
+    def __init__(
+        self,
+        block_frames=256,
+        max_frames=1024,
+        use_onnx=True,
+        min_frames=2,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.block_frames = block_frames
+        self.max_frames = max_frames
+        self.min_frames = min_frames
+
+        # Initialize Silero VAD
+        self._init_vad(use_onnx)
+
+    def _init_vad(self, use_onnx=False):
+        """Initialize Silero VAD model"""
+        try:
+            torch.set_num_threads(1)
+            self.vad_model = load_silero_vad(onnx=use_onnx)
+            self.vad_iterator = VADIterator(self.vad_model, sampling_rate=16000)
+            self.logger.info("Silero VAD initialized successfully")
+
+        except Exception as e:
+            self.logger.error(f"Failed to initialize Silero VAD: {e}")
+            self.vad_model = None
+            self.vad_iterator = None
+
+    async def _chunk(self, data: av.AudioFrame) -> Optional[list[av.AudioFrame]]:
+        """Process audio frame and return chunk when ready"""
+        self.frames.append(data)
+
+        # Check for speech segments every 32 frames (~1 second)
+        if len(self.frames) >= 32 and len(self.frames) % 32 == 0:
+            return await self._process_block()
+
+        # Safety fallback - emit if we hit max frames
+        elif len(self.frames) >= self.max_frames:
+            self.logger.warning(
+                f"AudioChunkerSileroProcessor: Reached max frames ({self.max_frames}), "
+                f"emitting first {self.max_frames // 2} frames"
+            )
+            frames_to_emit = self.frames[: self.max_frames // 2]
+            self.frames = self.frames[self.max_frames // 2 :]
+            if len(frames_to_emit) >= self.min_frames:
+                return frames_to_emit
+            else:
+                self.logger.debug(
+                    f"Ignoring fallback segment with {len(frames_to_emit)} frames "
+                    f"(< {self.min_frames} minimum)"
+                )
+
+        return None
+
+    async def _process_block(self) -> Optional[list[av.AudioFrame]]:
+        # Need at least 32 frames for VAD detection (~1 second)
+        if len(self.frames) < 32 or self.vad_iterator is None:
+            return None
+
+        # Processing block with current buffer size
+        print(f"Processing block: {len(self.frames)} frames in buffer")
+
+        try:
+            # Convert frames to numpy array for VAD
+            audio_array = self._frames_to_numpy(self.frames)
+
+            if audio_array is None:
+                # Fallback: emit all frames if conversion failed
+                frames_to_emit = self.frames[:]
+                self.frames = []
+                if len(frames_to_emit) >= self.min_frames:
+                    return frames_to_emit
+                else:
+                    self.logger.debug(
+                        f"Ignoring conversion-failed segment with {len(frames_to_emit)} frames "
+                        f"(< {self.min_frames} minimum)"
+                    )
+                return None
+
+            # Find complete speech segments in the buffer
+            speech_end_frame = self._find_speech_segment_end(audio_array)
+
+            if speech_end_frame is None or speech_end_frame <= 0:
+                # No speech found but buffer is getting large
+                if len(self.frames) > 512:
+                    # Check if it's all silence and can be discarded
+                    # No speech segment found, buffer at {len(self.frames)} frames
+
+                    # Could emit silence or discard old frames here
+                    # For now, keep first 256 frames and discard older silence
+                    if len(self.frames) > 768:
+                        self.logger.debug(
+                            f"Discarding {len(self.frames) - 256} old frames (likely silence)"
+                        )
+                        self.frames = self.frames[-256:]
+                return None
+
+            # Calculate segment timing information
+            frames_to_emit = self.frames[:speech_end_frame]
+
+            # Get timing from av.AudioFrame
+            if frames_to_emit:
+                first_frame = frames_to_emit[0]
+                last_frame = frames_to_emit[-1]
+                sample_rate = first_frame.sample_rate
+
+                # Calculate duration
+                total_samples = sum(f.samples for f in frames_to_emit)
+                duration_seconds = total_samples / sample_rate if sample_rate > 0 else 0
+
+                # Get timestamps if available
+                start_time = (
+                    first_frame.pts * first_frame.time_base if first_frame.pts else 0
+                )
+                end_time = (
+                    last_frame.pts * last_frame.time_base if last_frame.pts else 0
+                )
+
+                # Convert to HH:MM:SS format for logging
+                def format_time(seconds):
+                    if not seconds:
+                        return "00:00:00"
+                    total_seconds = int(float(seconds))
+                    hours = total_seconds // 3600
+                    minutes = (total_seconds % 3600) // 60
+                    secs = total_seconds % 60
+                    return f"{hours:02d}:{minutes:02d}:{secs:02d}"
+
+                start_formatted = format_time(start_time)
+                end_formatted = format_time(end_time)
+
+                # Keep remaining frames for next processing
+                remaining_after = len(self.frames) - speech_end_frame
+
+                # Single structured log line
+                self.logger.info(
+                    "Speech segment found",
+                    start=start_formatted,
+                    end=end_formatted,
+                    frames=speech_end_frame,
+                    duration=round(duration_seconds, 2),
+                    buffer_before=len(self.frames),
+                    remaining=remaining_after,
+                )
+
+            # Keep remaining frames for next processing
+            self.frames = self.frames[speech_end_frame:]
+
+            # Filter out segments with too few frames
+            if len(frames_to_emit) >= self.min_frames:
+                return frames_to_emit
+            else:
+                self.logger.debug(
+                    f"Ignoring segment with {len(frames_to_emit)} frames "
+                    f"(< {self.min_frames} minimum)"
+                )
+
+        except Exception as e:
+            self.logger.error(f"Error in VAD processing: {e}")
+            # Fallback to simple chunking
+            if len(self.frames) >= self.block_frames:
+                frames_to_emit = self.frames[: self.block_frames]
+                self.frames = self.frames[self.block_frames :]
+                if len(frames_to_emit) >= self.min_frames:
+                    return frames_to_emit
+                else:
+                    self.logger.debug(
+                        f"Ignoring exception-fallback segment with {len(frames_to_emit)} frames "
+                        f"(< {self.min_frames} minimum)"
+                    )
+
+        return None
+
+    def _frames_to_numpy(self, frames: list[av.AudioFrame]) -> Optional[np.ndarray]:
+        """Convert av.AudioFrame list to numpy array for VAD processing"""
+        if not frames:
+            return None
+
+        try:
+            audio_data = []
+            for frame in frames:
+                frame_array = frame.to_ndarray()
+
+                if len(frame_array.shape) == 2:
+                    frame_array = frame_array.flatten()
+
+                audio_data.append(frame_array)
+
+            if not audio_data:
+                return None
+
+            combined_audio = np.concatenate(audio_data)
+
+            # Ensure float32 format
+            if combined_audio.dtype == np.int16:
+                # Normalize int16 audio to float32 in range [-1.0, 1.0]
+                combined_audio = combined_audio.astype(np.float32) / 32768.0
+            elif combined_audio.dtype != np.float32:
+                combined_audio = combined_audio.astype(np.float32)
+
+            return combined_audio
+
+        except Exception as e:
+            self.logger.error(f"Error converting frames to numpy: {e}")
+
+        return None
+
+    def _find_speech_segment_end(self, audio_array: np.ndarray) -> Optional[int]:
+        """Find complete speech segments and return frame index at segment end"""
+        if self.vad_iterator is None or len(audio_array) == 0:
+            return None
+
+        try:
+            # Process audio in 512-sample windows for VAD
+            window_size = 512
+            min_silence_windows = 3  # Require 3 windows of silence after speech
+
+            # Track speech state
+            in_speech = False
+            speech_start = None
+            speech_end = None
+            silence_count = 0
+
+            for i in range(0, len(audio_array), window_size):
+                chunk = audio_array[i : i + window_size]
+                if len(chunk) < window_size:
+                    chunk = np.pad(chunk, (0, window_size - len(chunk)))
+
+                # Detect if this window has speech
+                speech_dict = self.vad_iterator(chunk, return_seconds=True)
+
+                # VADIterator returns dict with 'start' and 'end' when speech segments are detected
+                if speech_dict:
+                    if not in_speech:
+                        # Speech started
+                        speech_start = i
+                        in_speech = True
+                        # Debug: print(f"Speech START at sample {i}, VAD: {speech_dict}")
+                    silence_count = 0  # Reset silence counter
+                    continue
+
+                if not in_speech:
+                    continue
+
+                # We're in speech but found silence
+                silence_count += 1
+                if silence_count < min_silence_windows:
+                    continue
+
+                # Found end of speech segment
+                speech_end = i - (min_silence_windows - 1) * window_size
+                # Debug: print(f"Speech END at sample {speech_end}")
+
+                # Convert sample position to frame index
+                samples_per_frame = self.frames[0].samples if self.frames else 1024
+                frame_index = speech_end // samples_per_frame
+
+                # Ensure we don't exceed buffer
+                frame_index = min(frame_index, len(self.frames))
+                return frame_index
+
+            return None
+
+        except Exception as e:
+            self.logger.error(f"Error finding speech segment: {e}")
+            return None
+
+    async def _flush(self):
+        frames = self.frames[:]
+        self.frames = []
+        if frames:
+            if len(frames) >= self.min_frames:
+                await self.emit(frames)
+            else:
+                self.logger.debug(
+                    f"Ignoring flush segment with {len(frames)} frames "
+                    f"(< {self.min_frames} minimum)"
+                )
+
+
+AudioChunkerAutoProcessor.register("silero", AudioChunkerSileroProcessor)

server/reflector/processors/audio_downscale.py

@@ -0,0 +1,60 @@
+from typing import Optional
+
+import av
+from av.audio.resampler import AudioResampler
+
+from reflector.processors.base import Processor
+
+
+def copy_frame(frame: av.AudioFrame) -> av.AudioFrame:
+    frame_copy = frame.from_ndarray(
+        frame.to_ndarray(),
+        format=frame.format.name,
+        layout=frame.layout.name,
+    )
+    frame_copy.sample_rate = frame.sample_rate
+    frame_copy.pts = frame.pts
+    frame_copy.time_base = frame.time_base
+    return frame_copy
+
+
+class AudioDownscaleProcessor(Processor):
+    """
+    Downscale audio frames to 16kHz mono format
+    """
+
+    INPUT_TYPE = av.AudioFrame
+    OUTPUT_TYPE = av.AudioFrame
+
+    def __init__(self, target_rate: int = 16000, target_layout: str = "mono", **kwargs):
+        super().__init__(**kwargs)
+        self.target_rate = target_rate
+        self.target_layout = target_layout
+        self.resampler: Optional[AudioResampler] = None
+        self.needs_resampling: Optional[bool] = None
+
+    async def _push(self, data: av.AudioFrame):
+        if self.needs_resampling is None:
+            self.needs_resampling = (
+                data.sample_rate != self.target_rate
+                or data.layout.name != self.target_layout
+            )
+
+            if self.needs_resampling:
+                self.resampler = AudioResampler(
+                    format="s16", layout=self.target_layout, rate=self.target_rate
+                )
+
+        if not self.needs_resampling or not self.resampler:
+            await self.emit(data)
+            return
+
+        resampled_frames = self.resampler.resample(copy_frame(data))
+        for resampled_frame in resampled_frames:
+            await self.emit(resampled_frame)
+
+    async def _flush(self):
+        if self.needs_resampling and self.resampler:
+            final_frames = self.resampler.resample(None)
+            for frame in final_frames:
+                await self.emit(frame)

server/reflector/processors/audio_merge.py

@@ -3,24 +3,11 @@ from time import monotonic_ns
 from uuid import uuid4
 
 import av
-from av.audio.resampler import AudioResampler
 
 from reflector.processors.base import Processor
 from reflector.processors.types import AudioFile
 
 
-def copy_frame(frame: av.AudioFrame) -> av.AudioFrame:
-    frame_copy = frame.from_ndarray(
-        frame.to_ndarray(),
-        format=frame.format.name,
-        layout=frame.layout.name,
-    )
-    frame_copy.sample_rate = frame.sample_rate
-    frame_copy.pts = frame.pts
-    frame_copy.time_base = frame.time_base
-    return frame_copy
-
-
 class AudioMergeProcessor(Processor):
     """
     Merge audio frame into a single file
@@ -29,9 +16,8 @@ class AudioMergeProcessor(Processor):
     INPUT_TYPE = list[av.AudioFrame]
     OUTPUT_TYPE = AudioFile
 
-    def __init__(self, downsample_to_16k_mono: bool = True, **kwargs):
+    def __init__(self, **kwargs):
         super().__init__(**kwargs)
-        self.downsample_to_16k_mono = downsample_to_16k_mono
 
     async def _push(self, data: list[av.AudioFrame]):
         if not data:
@@ -39,72 +25,27 @@ class AudioMergeProcessor(Processor):
 
         # get audio information from first frame
         frame = data[0]
-        original_channels = len(frame.layout.channels)
-        original_sample_rate = frame.sample_rate
-        original_sample_width = frame.format.bytes
-
-        # determine if we need processing
-        needs_processing = self.downsample_to_16k_mono and (
-            original_sample_rate != 16000 or original_channels != 1
-        )
-
-        # determine output parameters
-        if self.downsample_to_16k_mono:
-            output_sample_rate = 16000
-            output_channels = 1
-            output_sample_width = 2  # 16-bit = 2 bytes
-        else:
-            output_sample_rate = original_sample_rate
-            output_channels = original_channels
-            output_sample_width = original_sample_width
+        output_channels = len(frame.layout.channels)
+        output_sample_rate = frame.sample_rate
+        output_sample_width = frame.format.bytes
 
         # create audio file
         uu = uuid4().hex
         fd = io.BytesIO()
 
-        if needs_processing:
-            # Process with PyAV resampler
-            out_container = av.open(fd, "w", format="wav")
-            out_stream = out_container.add_stream("pcm_s16le", rate=16000)
-            out_stream.layout = "mono"
+        # Use PyAV to write frames
+        out_container = av.open(fd, "w", format="wav")
+        out_stream = out_container.add_stream("pcm_s16le", rate=output_sample_rate)
+        out_stream.layout = frame.layout.name
 
-            # Create resampler if needed
-            resampler = None
-            if original_sample_rate != 16000 or original_channels != 1:
-                resampler = AudioResampler(format="s16", layout="mono", rate=16000)
-
-            for frame in data:
-                if resampler:
-                    # Resample and convert to mono
-                    # XXX for an unknown reason, if we don't use a copy of the frame, we get
-                    # Invalid Argumment from resample. Debugging indicate that when a previous processor
-                    # already used the frame (like AudioFileWriter), it make it invalid argument here.
-                    resampled_frames = resampler.resample(copy_frame(frame))
-                    for resampled_frame in resampled_frames:
-                        for packet in out_stream.encode(resampled_frame):
-                            out_container.mux(packet)
-                else:
-                    # Direct encoding without resampling
-                    for packet in out_stream.encode(frame):
-                        out_container.mux(packet)
-
-            # Flush the encoder
-            for packet in out_stream.encode(None):
+        for frame in data:
+            for packet in out_stream.encode(frame):
                 out_container.mux(packet)
-            out_container.close()
-        else:
-            # Use PyAV for original frames (no processing needed)
-            out_container = av.open(fd, "w", format="wav")
-            out_stream = out_container.add_stream("pcm_s16le", rate=output_sample_rate)
-            out_stream.layout = "mono" if output_channels == 1 else frame.layout
 
-            for frame in data:
-                for packet in out_stream.encode(frame):
-                    out_container.mux(packet)
-
-            for packet in out_stream.encode(None):
-                out_container.mux(packet)
-            out_container.close()
+        # Flush the encoder
+        for packet in out_stream.encode(None):
+            out_container.mux(packet)
+        out_container.close()
 
         fd.seek(0)
 

server/reflector/processors/audio_transcript_modal.py

@@ -12,9 +12,6 @@ API will be a POST request to TRANSCRIPT_URL:
 
 """
 
-from typing import List
-
-import aiohttp
 from openai import AsyncOpenAI
 
 from reflector.processors.audio_transcript import AudioTranscriptProcessor
@@ -25,7 +22,9 @@ from reflector.settings import settings
 
 class AudioTranscriptModalProcessor(AudioTranscriptProcessor):
     def __init__(
-        self, modal_api_key: str | None = None, batch_enabled: bool = True, **kwargs
+        self,
+        modal_api_key: str | None = None,
+        **kwargs,
     ):
         super().__init__()
         if not settings.TRANSCRIPT_URL:
@@ -35,126 +34,6 @@ class AudioTranscriptModalProcessor(AudioTranscriptProcessor):
         self.transcript_url = settings.TRANSCRIPT_URL + "/v1"
         self.timeout = settings.TRANSCRIPT_TIMEOUT
         self.modal_api_key = modal_api_key
-        self.max_batch_duration = 10.0
-        self.max_batch_files = 15
-        self.batch_enabled = batch_enabled
-        self.pending_files: List[AudioFile] = []  # Files waiting to be processed
-
-    @classmethod
-    def _calculate_duration(cls, audio_file: AudioFile) -> float:
-        """Calculate audio duration in seconds from AudioFile metadata"""
-        # Duration = total_samples / sample_rate
-        # We need to estimate total samples from the file data
-        import wave
-
-        try:
-            # Try to read as WAV file to get duration
-            audio_file.fd.seek(0)
-            with wave.open(audio_file.fd, "rb") as wav_file:
-                frames = wav_file.getnframes()
-                sample_rate = wav_file.getframerate()
-                duration = frames / sample_rate
-                return duration
-        except Exception:
-            # Fallback: estimate from file size and audio parameters
-            audio_file.fd.seek(0, 2)  # Seek to end
-            file_size = audio_file.fd.tell()
-            audio_file.fd.seek(0)  # Reset to beginning
-
-            # Estimate: file_size / (sample_rate * channels * sample_width)
-            bytes_per_second = (
-                audio_file.sample_rate
-                * audio_file.channels
-                * (audio_file.sample_width // 8)
-            )
-            estimated_duration = (
-                file_size / bytes_per_second if bytes_per_second > 0 else 0
-            )
-            return max(0, estimated_duration)
-
-    def _create_batches(self, audio_files: List[AudioFile]) -> List[List[AudioFile]]:
-        """Group audio files into batches with maximum 30s total duration"""
-        batches = []
-        current_batch = []
-        current_duration = 0.0
-
-        for audio_file in audio_files:
-            duration = self._calculate_duration(audio_file)
-
-            # If adding this file exceeds max duration, start a new batch
-            if current_duration + duration > self.max_batch_duration and current_batch:
-                batches.append(current_batch)
-                current_batch = [audio_file]
-                current_duration = duration
-            else:
-                current_batch.append(audio_file)
-                current_duration += duration
-
-        # Add the last batch if not empty
-        if current_batch:
-            batches.append(current_batch)
-
-        return batches
-
-    async def _transcript_batch(self, audio_files: List[AudioFile]) -> List[Transcript]:
-        """Transcribe a batch of audio files using the parakeet backend"""
-        if not audio_files:
-            return []
-
-        self.logger.debug(f"Batch transcribing {len(audio_files)} files")
-
-        # Prepare form data for batch request
-        data = aiohttp.FormData()
-        data.add_field("language", self.get_pref("audio:source_language", "en"))
-        data.add_field("batch", "true")
-
-        for i, audio_file in enumerate(audio_files):
-            audio_file.fd.seek(0)
-            data.add_field(
-                "files",
-                audio_file.fd,
-                filename=f"{audio_file.name}",
-                content_type="audio/wav",
-            )
-
-        # Make batch request
-        headers = {"Authorization": f"Bearer {self.modal_api_key}"}
-
-        async with aiohttp.ClientSession(
-            timeout=aiohttp.ClientTimeout(total=self.timeout)
-        ) as session:
-            async with session.post(
-                f"{self.transcript_url}/audio/transcriptions",
-                data=data,
-                headers=headers,
-            ) as response:
-                if response.status != 200:
-                    error_text = await response.text()
-                    raise Exception(
-                        f"Batch transcription failed: {response.status} {error_text}"
-                    )
-
-                result = await response.json()
-
-        # Process batch results
-        transcripts = []
-        results = result.get("results", [])
-
-        for i, (audio_file, file_result) in enumerate(zip(audio_files, results)):
-            transcript = Transcript(
-                words=[
-                    Word(
-                        text=word_info["word"],
-                        start=word_info["start"],
-                        end=word_info["end"],
-                    )
-                    for word_info in file_result.get("words", [])
-                ]
-            )
-            transcript.add_offset(audio_file.timestamp)
-            transcripts.append(transcript)
-
-        return transcripts
 
     async def _transcript(self, data: AudioFile):
         async with AsyncOpenAI(
@@ -187,96 +66,5 @@ class AudioTranscriptModalProcessor(AudioTranscriptProcessor):
 
         return transcript
 
-    async def transcript_multiple(
-        self, audio_files: List[AudioFile]
-    ) -> List[Transcript]:
-        """Transcribe multiple audio files using batching"""
-        if len(audio_files) == 1:
-            # Single file, use existing method
-            return [await self._transcript(audio_files[0])]
-
-        # Create batches with max 30s duration each
-        batches = self._create_batches(audio_files)
-
-        self.logger.debug(
-            f"Processing {len(audio_files)} files in {len(batches)} batches"
-        )
-
-        # Process all batches concurrently
-        all_transcripts = []
-
-        for batch in batches:
-            batch_transcripts = await self._transcript_batch(batch)
-            all_transcripts.extend(batch_transcripts)
-
-        return all_transcripts
-
-    async def _push(self, data: AudioFile):
-        """Override _push to support batching"""
-        if not self.batch_enabled:
-            # Use parent implementation for single file processing
-            return await super()._push(data)
-
-        # Add file to pending batch
-        self.pending_files.append(data)
-        self.logger.debug(
-            f"Added file to batch: {data.name}, batch size: {len(self.pending_files)}"
-        )
-
-        # Calculate total duration of pending files
-        total_duration = sum(self._calculate_duration(f) for f in self.pending_files)
-
-        # Process batch if it reaches max duration or has multiple files ready for optimization
-        should_process_batch = (
-            total_duration >= self.max_batch_duration
-            or len(self.pending_files) >= self.max_batch_files
-        )
-
-        if should_process_batch:
-            await self._process_pending_batch()
-
-    async def _process_pending_batch(self):
-        """Process all pending files as batches"""
-        if not self.pending_files:
-            return
-
-        self.logger.debug(f"Processing batch of {len(self.pending_files)} files")
-
-        try:
-            # Create batches respecting duration limit
-            batches = self._create_batches(self.pending_files)
-
-            # Process each batch
-            for batch in batches:
-                self.m_transcript_call.inc()
-                try:
-                    with self.m_transcript.time():
-                        # Use batch transcription
-                        transcripts = await self._transcript_batch(batch)
-
-                    self.m_transcript_success.inc()
-
-                    # Emit each transcript
-                    for transcript in transcripts:
-                        if transcript:
-                            await self.emit(transcript)
-
-                except Exception:
-                    self.m_transcript_failure.inc()
-                    raise
-                finally:
-                    # Release audio files
-                    for audio_file in batch:
-                        audio_file.release()
-
-        finally:
-            # Clear pending files
-            self.pending_files.clear()
-
-    async def _flush(self):
-        """Process any remaining files when flushing"""
-        await self._process_pending_batch()
-        await super()._flush()
-
 
 AudioTranscriptAutoProcessor.register("modal", AudioTranscriptModalProcessor)

server/reflector/processors/file_diarization_modal.py

@@ -47,6 +47,7 @@ class FileDiarizationModalProcessor(FileDiarizationProcessor):
                     "audio_file_url": data.audio_url,
                     "timestamp": 0,
                 },
+                follow_redirects=True,
             )
             response.raise_for_status()
             diarization_data = response.json()["diarization"]

server/reflector/processors/file_transcript_modal.py

@@ -54,6 +54,7 @@ class FileTranscriptModalProcessor(FileTranscriptProcessor):
                     "language": data.language,
                     "batch": True,
                 },
+                follow_redirects=True,
             )
             response.raise_for_status()
             result = response.json()
@@ -67,6 +68,9 @@ class FileTranscriptModalProcessor(FileTranscriptProcessor):
             for word_info in result.get("words", [])
         ]
 
+        # words come not in order
+        words.sort(key=lambda w: w.start)
+
         return Transcript(words=words)
 
 

server/reflector/processors/transcript_topic_detector.py

@@ -1,6 +1,6 @@
 from textwrap import dedent
 
-from pydantic import BaseModel, Field
+from pydantic import BaseModel, ConfigDict, Field
 
 from reflector.llm import LLM
 from reflector.processors.base import Processor
@@ -34,8 +34,14 @@ TOPIC_PROMPT = dedent(
 class TopicResponse(BaseModel):
     """Structured response for topic detection"""
 
-    title: str = Field(description="A descriptive title for the topic being discussed")
-    summary: str = Field(description="A concise 1-2 sentence summary of the discussion")
+    model_config = ConfigDict(populate_by_name=True)
+
+    title: str = Field(
+        description="A descriptive title for the topic being discussed", alias="Title"
+    )
+    summary: str = Field(
+        description="A concise 1-2 sentence summary of the discussion", alias="Summary"
+    )
 
 
 class TranscriptTopicDetectorProcessor(Processor):

server/reflector/processors/types.py

@@ -4,11 +4,8 @@ import tempfile
 from pathlib import Path
 from typing import Annotated, TypedDict
 
-from profanityfilter import ProfanityFilter
 from pydantic import BaseModel, Field, PrivateAttr
 
-from reflector.redis_cache import redis_cache
-
 
 class DiarizationSegment(TypedDict):
     """Type definition for diarization segment containing speaker information"""
@@ -20,9 +17,6 @@ class DiarizationSegment(TypedDict):
 
 PUNC_RE = re.compile(r"[.;:?!…]")
 
-profanity_filter = ProfanityFilter()
-profanity_filter.set_censor("*")
-
 
 class AudioFile(BaseModel):
     name: str
@@ -124,21 +118,11 @@ def words_to_segments(words: list[Word]) -> list[TranscriptSegment]:
 
 class Transcript(BaseModel):
     translation: str | None = None
-    words: list[Word] = None
-
-    @property
-    def raw_text(self):
-        # Uncensored text
-        return "".join([word.text for word in self.words])
-
-    @redis_cache(prefix="profanity", duration=3600 * 24 * 7)
-    def _get_censored_text(self, text: str):
-        return profanity_filter.censor(text).strip()
+    words: list[Word] = []
 
     @property
     def text(self):
-        # Censored text
-        return self._get_censored_text(self.raw_text)
+        return "".join([word.text for word in self.words])
 
     @property
     def human_timestamp(self):
@@ -170,12 +154,6 @@ class Transcript(BaseModel):
             word.start += offset
             word.end += offset
 
-    def clone(self):
-        words = [
-            Word(text=word.text, start=word.start, end=word.end) for word in self.words
-        ]
-        return Transcript(text=self.text, translation=self.translation, words=words)
-
     def as_segments(self) -> list[TranscriptSegment]:
         return words_to_segments(self.words)
 

server/reflector/redis_cache.py

@@ -1,10 +1,17 @@
+import asyncio
 import functools
 import json
+from typing import Optional
 
 import redis
+import redis.asyncio as redis_async
+import structlog
+from redis.exceptions import LockError
 
 from reflector.settings import settings
 
+logger = structlog.get_logger(__name__)
+
 redis_clients = {}
 
 
@@ -21,6 +28,12 @@ def get_redis_client(db=0):
     return redis_clients[db]
 
 
+async def get_async_redis_client(db: int = 0):
+    return await redis_async.from_url(
+        f"redis://{settings.REDIS_HOST}:{settings.REDIS_PORT}/{db}"
+    )
+
+
 def redis_cache(prefix="cache", duration=3600, db=settings.REDIS_CACHE_DB, argidx=1):
     """
     Cache the result of a function in Redis.
@@ -49,3 +62,87 @@ def redis_cache(prefix="cache", duration=3600, db=settings.REDIS_CACHE_DB, argid
         return wrapper
 
     return decorator
+
+
+class RedisAsyncLock:
+    def __init__(
+        self,
+        key: str,
+        timeout: int = 120,
+        extend_interval: int = 30,
+        skip_if_locked: bool = False,
+        blocking: bool = True,
+        blocking_timeout: Optional[float] = None,
+    ):
+        self.key = f"async_lock:{key}"
+        self.timeout = timeout
+        self.extend_interval = extend_interval
+        self.skip_if_locked = skip_if_locked
+        self.blocking = blocking
+        self.blocking_timeout = blocking_timeout
+        self._lock = None
+        self._redis = None
+        self._extend_task = None
+        self._acquired = False
+
+    async def _extend_lock_periodically(self):
+        while True:
+            try:
+                await asyncio.sleep(self.extend_interval)
+                if self._lock:
+                    await self._lock.extend(self.timeout, replace_ttl=True)
+                    logger.debug("Extended lock", key=self.key)
+            except LockError:
+                logger.warning("Failed to extend lock", key=self.key)
+                break
+            except asyncio.CancelledError:
+                break
+            except Exception as e:
+                logger.error("Error extending lock", key=self.key, error=str(e))
+                break
+
+    async def __aenter__(self):
+        self._redis = await get_async_redis_client()
+        self._lock = self._redis.lock(
+            self.key,
+            timeout=self.timeout,
+            blocking=self.blocking,
+            blocking_timeout=self.blocking_timeout,
+        )
+
+        self._acquired = await self._lock.acquire()
+
+        if not self._acquired:
+            if self.skip_if_locked:
+                logger.warning(
+                    "Lock already acquired by another process, skipping", key=self.key
+                )
+                return self
+            else:
+                raise LockError(f"Failed to acquire lock: {self.key}")
+
+        self._extend_task = asyncio.create_task(self._extend_lock_periodically())
+        logger.info("Acquired lock", key=self.key)
+        return self
+
+    async def __aexit__(self, exc_type, exc_val, exc_tb):
+        if self._extend_task:
+            self._extend_task.cancel()
+            try:
+                await self._extend_task
+            except asyncio.CancelledError:
+                pass
+
+        if self._acquired and self._lock:
+            try:
+                await self._lock.release()
+                logger.info("Released lock", key=self.key)
+            except LockError:
+                logger.debug("Lock already released or expired", key=self.key)
+
+        if self._redis:
+            await self._redis.aclose()
+
+    @property
+    def acquired(self) -> bool:
+        return self._acquired

server/reflector/services/ics_sync.py

@@ -0,0 +1,408 @@
+"""
+ICS Calendar Synchronization Service
+
+This module provides services for fetching, parsing, and synchronizing ICS (iCalendar)
+calendar feeds with room booking data in the database.
+
+Key Components:
+- ICSFetchService: Handles HTTP fetching and parsing of ICS calendar data
+- ICSSyncService: Manages the synchronization process between ICS feeds and database
+
+Example Usage:
+    # Sync a room's calendar
+    room = Room(id="room1", name="conference-room", ics_url="https://cal.example.com/room.ics")
+    result = await ics_sync_service.sync_room_calendar(room)
+
+    # Result structure:
+    {
+        "status": "success",  # success|unchanged|error|skipped
+        "hash": "abc123...",  # MD5 hash of ICS content
+        "events_found": 5,    # Events matching this room
+        "total_events": 12,   # Total events in calendar within time window
+        "events_created": 2,  # New events added to database
+        "events_updated": 3,  # Existing events modified
+        "events_deleted": 1   # Events soft-deleted (no longer in calendar)
+    }
+
+Event Matching:
+    Events are matched to rooms by checking if the room's full URL appears in the
+    event's LOCATION or DESCRIPTION fields. Only events within a 25-hour window
+    (1 hour ago to 24 hours from now) are processed.
+
+Input: ICS calendar URL (e.g., "https://calendar.google.com/calendar/ical/...")
+Output: EventData objects with structured calendar information:
+    {
+        "ics_uid": "event123@google.com",
+        "title": "Team Meeting",
+        "description": "Weekly sync meeting",
+        "location": "https://meet.company.com/conference-room",
+        "start_time": datetime(2024, 1, 15, 14, 0, tzinfo=UTC),
+        "end_time": datetime(2024, 1, 15, 15, 0, tzinfo=UTC),
+        "attendees": [
+            {"email": "user@company.com", "name": "John Doe", "role": "ORGANIZER"},
+            {"email": "attendee@company.com", "name": "Jane Smith", "status": "ACCEPTED"}
+        ],
+        "ics_raw_data": "BEGIN:VEVENT\nUID:event123@google.com\n..."
+    }
+"""
+
+import hashlib
+from datetime import date, datetime, timedelta, timezone
+from enum import Enum
+from typing import TypedDict
+
+import httpx
+import pytz
+import structlog
+from icalendar import Calendar, Event
+
+from reflector.db.calendar_events import CalendarEvent, calendar_events_controller
+from reflector.db.rooms import Room, rooms_controller
+from reflector.redis_cache import RedisAsyncLock
+from reflector.settings import settings
+
+logger = structlog.get_logger()
+
+EVENT_WINDOW_DELTA_START = timedelta(hours=-1)
+EVENT_WINDOW_DELTA_END = timedelta(hours=24)
+
+
+class SyncStatus(str, Enum):
+    SUCCESS = "success"
+    UNCHANGED = "unchanged"
+    ERROR = "error"
+    SKIPPED = "skipped"
+
+
+class AttendeeData(TypedDict, total=False):
+    email: str | None
+    name: str | None
+    status: str | None
+    role: str | None
+
+
+class EventData(TypedDict):
+    ics_uid: str
+    title: str | None
+    description: str | None
+    location: str | None
+    start_time: datetime
+    end_time: datetime
+    attendees: list[AttendeeData]
+    ics_raw_data: str
+
+
+class SyncStats(TypedDict):
+    events_created: int
+    events_updated: int
+    events_deleted: int
+
+
+class SyncResultBase(TypedDict):
+    status: SyncStatus
+
+
+class SyncResult(SyncResultBase, total=False):
+    hash: str | None
+    events_found: int
+    total_events: int
+    events_created: int
+    events_updated: int
+    events_deleted: int
+    error: str | None
+    reason: str | None
+
+
+class ICSFetchService:
+    def __init__(self):
+        self.client = httpx.AsyncClient(
+            timeout=30.0, headers={"User-Agent": "Reflector/1.0"}
+        )
+
+    async def fetch_ics(self, url: str) -> str:
+        response = await self.client.get(url)
+        response.raise_for_status()
+
+        return response.text
+
+    def parse_ics(self, ics_content: str) -> Calendar:
+        return Calendar.from_ical(ics_content)
+
+    def extract_room_events(
+        self, calendar: Calendar, room_name: str, room_url: str
+    ) -> tuple[list[EventData], int]:
+        events = []
+        total_events = 0
+        now = datetime.now(timezone.utc)
+        window_start = now + EVENT_WINDOW_DELTA_START
+        window_end = now + EVENT_WINDOW_DELTA_END
+
+        for component in calendar.walk():
+            if component.name != "VEVENT":
+                continue
+
+            status = component.get("STATUS", "").upper()
+            if status == "CANCELLED":
+                continue
+
+            # Count total non-cancelled events in the time window
+            event_data = self._parse_event(component)
+            if event_data and window_start <= event_data["start_time"] <= window_end:
+                total_events += 1
+
+                # Check if event matches this room
+                if self._event_matches_room(component, room_name, room_url):
+                    events.append(event_data)
+
+        return events, total_events
+
+    def _event_matches_room(self, event: Event, room_name: str, room_url: str) -> bool:
+        location = str(event.get("LOCATION", ""))
+        description = str(event.get("DESCRIPTION", ""))
+
+        # Only match full room URL
+        # XXX leaved here as a patterns, to later be extended with tinyurl or such too
+        patterns = [
+            room_url,
+        ]
+
+        # Check location and description for patterns
+        text_to_check = f"{location} {description}".lower()
+        for pattern in patterns:
+            if pattern.lower() in text_to_check:
+                return True
+
+        return False
+
+    def _parse_event(self, event: Event) -> EventData | None:
+        uid = str(event.get("UID", ""))
+        summary = str(event.get("SUMMARY", ""))
+        description = str(event.get("DESCRIPTION", ""))
+        location = str(event.get("LOCATION", ""))
+        dtstart = event.get("DTSTART")
+        dtend = event.get("DTEND")
+
+        if not dtstart:
+            return None
+
+        # Convert fields
+        start_time = self._normalize_datetime(
+            dtstart.dt if hasattr(dtstart, "dt") else dtstart
+        )
+        end_time = (
+            self._normalize_datetime(dtend.dt if hasattr(dtend, "dt") else dtend)
+            if dtend
+            else start_time + timedelta(hours=1)
+        )
+        attendees = self._parse_attendees(event)
+
+        # Get raw event data for storage
+        raw_data = event.to_ical().decode("utf-8")
+
+        return {
+            "ics_uid": uid,
+            "title": summary,
+            "description": description,
+            "location": location,
+            "start_time": start_time,
+            "end_time": end_time,
+            "attendees": attendees,
+            "ics_raw_data": raw_data,
+        }
+
+    def _normalize_datetime(self, dt) -> datetime:
+        # Ensure datetime is with timezone, if not, assume UTC
+        if isinstance(dt, date) and not isinstance(dt, datetime):
+            dt = datetime.combine(dt, datetime.min.time())
+            dt = pytz.UTC.localize(dt)
+        elif isinstance(dt, datetime):
+            if dt.tzinfo is None:
+                dt = pytz.UTC.localize(dt)
+            else:
+                dt = dt.astimezone(pytz.UTC)
+
+        return dt
+
+    def _parse_attendees(self, event: Event) -> list[AttendeeData]:
+        # Extracts attendee information from both ATTENDEE and ORGANIZER properties.
+        # Handles malformed comma-separated email addresses in single ATTENDEE fields
+        # by splitting them into separate attendee entries. Returns a list of attendee
+        # data including email, name, status, and role information.
+        final_attendees = []
+
+        attendees = event.get("ATTENDEE", [])
+        if not isinstance(attendees, list):
+            attendees = [attendees]
+        for att in attendees:
+            email_str = str(att).replace("mailto:", "") if att else None
+
+            # Handle malformed comma-separated email addresses in a single ATTENDEE field
+            if email_str and "," in email_str:
+                # Split comma-separated emails and create separate attendee entries
+                email_parts = [email.strip() for email in email_str.split(",")]
+                for email in email_parts:
+                    if email and "@" in email:
+                        clean_email = email.replace("MAILTO:", "").replace(
+                            "mailto:", ""
+                        )
+                        att_data: AttendeeData = {
+                            "email": clean_email,
+                            "name": att.params.get("CN")
+                            if hasattr(att, "params") and email == email_parts[0]
+                            else None,
+                            "status": att.params.get("PARTSTAT")
+                            if hasattr(att, "params") and email == email_parts[0]
+                            else None,
+                            "role": att.params.get("ROLE")
+                            if hasattr(att, "params") and email == email_parts[0]
+                            else None,
+                        }
+                        final_attendees.append(att_data)
+            else:
+                # Normal single attendee
+                att_data: AttendeeData = {
+                    "email": email_str,
+                    "name": att.params.get("CN") if hasattr(att, "params") else None,
+                    "status": att.params.get("PARTSTAT")
+                    if hasattr(att, "params")
+                    else None,
+                    "role": att.params.get("ROLE") if hasattr(att, "params") else None,
+                }
+                final_attendees.append(att_data)
+
+        # Add organizer
+        organizer = event.get("ORGANIZER")
+        if organizer:
+            org_email = (
+                str(organizer).replace("mailto:", "").replace("MAILTO:", "")
+                if organizer
+                else None
+            )
+            org_data: AttendeeData = {
+                "email": org_email,
+                "name": organizer.params.get("CN")
+                if hasattr(organizer, "params")
+                else None,
+                "role": "ORGANIZER",
+            }
+            final_attendees.append(org_data)
+
+        return final_attendees
+
+
+class ICSSyncService:
+    def __init__(self):
+        self.fetch_service = ICSFetchService()
+
+    async def sync_room_calendar(self, room: Room) -> SyncResult:
+        async with RedisAsyncLock(
+            f"ics_sync_room:{room.id}", skip_if_locked=True
+        ) as lock:
+            if not lock.acquired:
+                logger.warning("ICS sync already in progress for room", room_id=room.id)
+                return {
+                    "status": SyncStatus.SKIPPED,
+                    "reason": "Sync already in progress",
+                }
+
+            return await self._sync_room_calendar(room)
+
+    async def _sync_room_calendar(self, room: Room) -> SyncResult:
+        if not room.ics_enabled or not room.ics_url:
+            return {"status": SyncStatus.SKIPPED, "reason": "ICS not configured"}
+
+        try:
+            if not self._should_sync(room):
+                return {"status": SyncStatus.SKIPPED, "reason": "Not time to sync yet"}
+
+            ics_content = await self.fetch_service.fetch_ics(room.ics_url)
+            calendar = self.fetch_service.parse_ics(ics_content)
+
+            content_hash = hashlib.md5(ics_content.encode()).hexdigest()
+            if room.ics_last_etag == content_hash:
+                logger.info("No changes in ICS for room", room_id=room.id)
+                room_url = f"{settings.UI_BASE_URL}/{room.name}"
+                events, total_events = self.fetch_service.extract_room_events(
+                    calendar, room.name, room_url
+                )
+                return {
+                    "status": SyncStatus.UNCHANGED,
+                    "hash": content_hash,
+                    "events_found": len(events),
+                    "total_events": total_events,
+                    "events_created": 0,
+                    "events_updated": 0,
+                    "events_deleted": 0,
+                }
+
+            # Extract matching events
+            room_url = f"{settings.UI_BASE_URL}/{room.name}"
+            events, total_events = self.fetch_service.extract_room_events(
+                calendar, room.name, room_url
+            )
+            sync_result = await self._sync_events_to_database(room.id, events)
+
+            # Update room sync metadata
+            await rooms_controller.update(
+                room,
+                {
+                    "ics_last_sync": datetime.now(timezone.utc),
+                    "ics_last_etag": content_hash,
+                },
+                mutate=False,
+            )
+
+            return {
+                "status": SyncStatus.SUCCESS,
+                "hash": content_hash,
+                "events_found": len(events),
+                "total_events": total_events,
+                **sync_result,
+            }
+
+        except Exception as e:
+            logger.error("Failed to sync ICS for room", room_id=room.id, error=str(e))
+            return {"status": SyncStatus.ERROR, "error": str(e)}
+
+    def _should_sync(self, room: Room) -> bool:
+        if not room.ics_last_sync:
+            return True
+
+        time_since_sync = datetime.now(timezone.utc) - room.ics_last_sync
+        return time_since_sync.total_seconds() >= room.ics_fetch_interval
+
+    async def _sync_events_to_database(
+        self, room_id: str, events: list[EventData]
+    ) -> SyncStats:
+        created = 0
+        updated = 0
+
+        current_ics_uids = []
+
+        for event_data in events:
+            calendar_event = CalendarEvent(room_id=room_id, **event_data)
+            existing = await calendar_events_controller.get_by_ics_uid(
+                room_id, event_data["ics_uid"]
+            )
+
+            if existing:
+                updated += 1
+            else:
+                created += 1
+
+            await calendar_events_controller.upsert(calendar_event)
+            current_ics_uids.append(event_data["ics_uid"])
+
+        # Soft delete events that are no longer in calendar
+        deleted = await calendar_events_controller.soft_delete_missing(
+            room_id, current_ics_uids
+        )
+
+        return {
+            "events_created": created,
+            "events_updated": updated,
+            "events_deleted": deleted,
+        }
+
+
+ics_sync_service = ICSSyncService()

server/reflector/settings.py

@@ -1,5 +1,9 @@
+from pydantic.types import PositiveInt
 from pydantic_settings import BaseSettings, SettingsConfigDict
 
+from reflector.platform_types import Platform
+from reflector.utils.string import NonEmptyString
+
 
 class Settings(BaseSettings):
     model_config = SettingsConfigDict(
@@ -21,6 +25,10 @@ class Settings(BaseSettings):
     # local data directory
     DATA_DIR: str = "./data"
 
+    # Audio Chunking
+    # backends: silero, frames
+    AUDIO_CHUNKER_BACKEND: str = "frames"
+
     # Audio Transcription
     # backends: whisper, modal
     TRANSCRIPT_BACKEND: str = "whisper"
@@ -86,9 +94,8 @@ class Settings(BaseSettings):
     AUTH_JWT_PUBLIC_KEY: str | None = "authentik.monadical.com_public.pem"
     AUTH_JWT_AUDIENCE: str | None = None
 
-    # API public mode
-    # if set, all anonymous record will be public
     PUBLIC_MODE: bool = False
+    PUBLIC_DATA_RETENTION_DAYS: PositiveInt = 7
 
     # Min transcript length to generate topic + summary
     MIN_TRANSCRIPT_LENGTH: int = 750
@@ -116,13 +123,26 @@ class Settings(BaseSettings):
 
     # Whereby integration
     WHEREBY_API_URL: str = "https://api.whereby.dev/v1"
-    WHEREBY_API_KEY: str | None = None
+    WHEREBY_API_KEY: NonEmptyString | None = None
     WHEREBY_WEBHOOK_SECRET: str | None = None
     AWS_WHEREBY_ACCESS_KEY_ID: str | None = None
     AWS_WHEREBY_ACCESS_KEY_SECRET: str | None = None
     AWS_PROCESS_RECORDING_QUEUE_URL: str | None = None
     SQS_POLLING_TIMEOUT_SECONDS: int = 60
 
+    # Daily.co integration
+    DAILY_API_KEY: str | None = None
+    DAILY_WEBHOOK_SECRET: str | None = None
+    DAILY_SUBDOMAIN: str | None = None
+    AWS_DAILY_S3_BUCKET: str | None = None
+    AWS_DAILY_S3_REGION: str = "us-west-2"
+    AWS_DAILY_ROLE_ARN: str | None = None
+
+    # Platform Migration Feature Flags
+    DAILY_MIGRATION_ENABLED: bool = False
+    DAILY_MIGRATION_ROOM_IDS: list[str] = []
+    DEFAULT_VIDEO_PLATFORM: Platform = "whereby"
+
     # Zulip integration
     ZULIP_REALM: str | None = None
     ZULIP_API_KEY: str | None = None

server/reflector/tools/cleanup_old_data.py

@@ -0,0 +1,72 @@
+#!/usr/bin/env python
+"""
+Manual cleanup tool for old public data.
+Uses the same implementation as the Celery worker task.
+"""
+
+import argparse
+import asyncio
+import sys
+
+import structlog
+
+from reflector.settings import settings
+from reflector.worker.cleanup import _cleanup_old_public_data
+
+logger = structlog.get_logger(__name__)
+
+
+async def cleanup_old_data(days: int = 7):
+    logger.info(
+        "Starting manual cleanup",
+        retention_days=days,
+        public_mode=settings.PUBLIC_MODE,
+    )
+
+    if not settings.PUBLIC_MODE:
+        logger.critical(
+            "WARNING: PUBLIC_MODE is False. "
+            "This tool is intended for public instances only."
+        )
+        raise Exception("Tool intended for public instances only")
+
+    result = await _cleanup_old_public_data(days=days)
+
+    if result:
+        logger.info(
+            "Cleanup completed",
+            transcripts_deleted=result.get("transcripts_deleted", 0),
+            meetings_deleted=result.get("meetings_deleted", 0),
+            recordings_deleted=result.get("recordings_deleted", 0),
+            errors_count=len(result.get("errors", [])),
+        )
+        if result.get("errors"):
+            logger.warning(
+                "Errors encountered during cleanup:", errors=result["errors"][:10]
+            )
+    else:
+        logger.info("Cleanup skipped or completed without results")
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Clean up old transcripts and meetings"
+    )
+    parser.add_argument(
+        "--days",
+        type=int,
+        default=7,
+        help="Number of days to keep data (default: 7)",
+    )
+
+    args = parser.parse_args()
+
+    if args.days < 1:
+        logger.error("Days must be at least 1")
+        sys.exit(1)
+
+    asyncio.run(cleanup_old_data(days=args.days))
+
+
+if __name__ == "__main__":
+    main()

server/reflector/tools/process.py

@@ -1,292 +1,204 @@
 """
 Process audio file with diarization support
-===========================================
-
-Extended version of process.py that includes speaker diarization.
-This tool processes audio files locally without requiring the full server infrastructure.
 """
 
+import argparse
 import asyncio
-import tempfile
-import uuid
+import json
+import shutil
+import sys
+import time
 from pathlib import Path
-from typing import List
-
-import av
+from typing import Any, Dict, List, Literal
 
+from reflector.db.transcripts import SourceKind, TranscriptTopic, transcripts_controller
 from reflector.logger import logger
-from reflector.processors import (
-    AudioChunkerProcessor,
-    AudioFileWriterProcessor,
-    AudioMergeProcessor,
-    AudioTranscriptAutoProcessor,
-    Pipeline,
-    PipelineEvent,
-    TranscriptFinalSummaryProcessor,
-    TranscriptFinalTitleProcessor,
-    TranscriptLinerProcessor,
-    TranscriptTopicDetectorProcessor,
-    TranscriptTranslatorAutoProcessor,
+from reflector.pipelines.main_file_pipeline import (
+    task_pipeline_file_process as task_pipeline_file_process,
 )
-from reflector.processors.base import BroadcastProcessor, Processor
-from reflector.processors.types import (
-    AudioDiarizationInput,
-    TitleSummary,
-    TitleSummaryWithId,
+from reflector.pipelines.main_live_pipeline import pipeline_post as live_pipeline_post
+from reflector.pipelines.main_live_pipeline import (
+    pipeline_process as live_pipeline_process,
 )
 
 
-class TopicCollectorProcessor(Processor):
-    """Collect topics for diarization"""
+def serialize_topics(topics: List[TranscriptTopic]) -> List[Dict[str, Any]]:
+    """Convert TranscriptTopic objects to JSON-serializable dicts"""
+    serialized = []
+    for topic in topics:
+        topic_dict = topic.model_dump()
+        serialized.append(topic_dict)
+    return serialized
 
-    INPUT_TYPE = TitleSummary
-    OUTPUT_TYPE = TitleSummary
 
-    def __init__(self, **kwargs):
-        super().__init__(**kwargs)
-        self.topics: List[TitleSummaryWithId] = []
-        self._topic_id = 0
+def debug_print_speakers(serialized_topics: List[Dict[str, Any]]) -> None:
+    """Print debug info about speakers found in topics"""
+    all_speakers = set()
+    for topic_dict in serialized_topics:
+        for word in topic_dict.get("words", []):
+            all_speakers.add(word.get("speaker", 0))
 
-    async def _push(self, data: TitleSummary):
-        # Convert to TitleSummaryWithId and collect
-        self._topic_id += 1
-        topic_with_id = TitleSummaryWithId(
-            id=str(self._topic_id),
-            title=data.title,
-            summary=data.summary,
-            timestamp=data.timestamp,
-            duration=data.duration,
-            transcript=data.transcript,
+    print(
+        f"Found {len(serialized_topics)} topics with speakers: {all_speakers}",
+        file=sys.stderr,
+    )
+
+
+TranscriptId = str
+
+
+# common interface for every flow: it needs an Entry in db with specific ceremony (file path + status + actual file in file system)
+# ideally we want to get rid of it at some point
+async def prepare_entry(
+    source_path: str,
+    source_language: str,
+    target_language: str,
+) -> TranscriptId:
+    file_path = Path(source_path)
+
+    transcript = await transcripts_controller.add(
+        file_path.name,
+        # note that the real file upload has SourceKind: LIVE for the reason of it's an error
+        source_kind=SourceKind.FILE,
+        source_language=source_language,
+        target_language=target_language,
+        user_id=None,
+    )
+
+    logger.info(
+        f"Created empty transcript {transcript.id} for file {file_path.name} because technically we need an empty transcript before we start transcript"
+    )
+
+    # pipelines expect files as upload.*
+
+    extension = file_path.suffix
+    upload_path = transcript.data_path / f"upload{extension}"
+    upload_path.parent.mkdir(parents=True, exist_ok=True)
+    shutil.copy2(source_path, upload_path)
+    logger.info(f"Copied {source_path} to {upload_path}")
+
+    # pipelines expect entity status "uploaded"
+    await transcripts_controller.update(transcript, {"status": "uploaded"})
+
+    return transcript.id
+
+
+# same reason as prepare_entry
+async def extract_result_from_entry(
+    transcript_id: TranscriptId, output_path: str
+) -> None:
+    post_final_transcript = await transcripts_controller.get_by_id(transcript_id)
+
+    # assert post_final_transcript.status == "ended"
+    # File pipeline doesn't set status to "ended", only live pipeline does https://github.com/Monadical-SAS/reflector/issues/582
+    topics = post_final_transcript.topics
+    if not topics:
+        raise RuntimeError(
+            f"No topics found for transcript {transcript_id} after processing"
         )
-        self.topics.append(topic_with_id)
 
-        # Pass through the original topic
-        await self.emit(data)
+    serialized_topics = serialize_topics(topics)
 
-    def get_topics(self) -> List[TitleSummaryWithId]:
-        return self.topics
+    if output_path:
+        # Write to JSON file
+        with open(output_path, "w") as f:
+            for topic_dict in serialized_topics:
+                json.dump(topic_dict, f)
+                f.write("\n")
+        print(f"Results written to {output_path}", file=sys.stderr)
+    else:
+        # Write to stdout as JSONL
+        for topic_dict in serialized_topics:
+            print(json.dumps(topic_dict))
+
+    debug_print_speakers(serialized_topics)
 
 
-async def process_audio_file(
-    filename,
-    event_callback,
-    only_transcript=False,
-    source_language="en",
-    target_language="en",
-    enable_diarization=True,
-    diarization_backend="pyannote",
+async def process_live_pipeline(
+    transcript_id: TranscriptId,
 ):
-    # Create temp file for audio if diarization is enabled
-    audio_temp_path = None
-    if enable_diarization:
-        audio_temp_file = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
-        audio_temp_path = audio_temp_file.name
-        audio_temp_file.close()
+    """Process transcript_id with transcription and diarization"""
 
-    # Create processor for collecting topics
-    topic_collector = TopicCollectorProcessor()
+    print(f"Processing transcript_id {transcript_id}...", file=sys.stderr)
+    await live_pipeline_process(transcript_id=transcript_id)
+    print(f"Processing complete for transcript {transcript_id}", file=sys.stderr)
 
-    # Build pipeline for audio processing
-    processors = []
+    pre_final_transcript = await transcripts_controller.get_by_id(transcript_id)
 
-    # Add audio file writer at the beginning if diarization is enabled
-    if enable_diarization:
-        processors.append(AudioFileWriterProcessor(audio_temp_path))
+    # assert documented behaviour: after process, the pipeline isn't ended. this is the reason of calling pipeline_post
+    assert pre_final_transcript.status != "ended"
 
-    # Add the rest of the processors
-    processors += [
-        AudioChunkerProcessor(),
-        AudioMergeProcessor(),
-        AudioTranscriptAutoProcessor.as_threaded(),
-        TranscriptLinerProcessor(),
-        TranscriptTranslatorAutoProcessor.as_threaded(),
-    ]
+    # at this point, diarization is running but we have no access to it. run diarization in parallel - one will hopefully win after polling
+    result = live_pipeline_post(transcript_id=transcript_id)
 
-    if not only_transcript:
-        processors += [
-            TranscriptTopicDetectorProcessor.as_threaded(),
-            # Collect topics for diarization
-            topic_collector,
-            BroadcastProcessor(
-                processors=[
-                    TranscriptFinalTitleProcessor.as_threaded(),
-                    TranscriptFinalSummaryProcessor.as_threaded(),
-                ],
-            ),
-        ]
-
-    # Create main pipeline
-    pipeline = Pipeline(*processors)
-    pipeline.set_pref("audio:source_language", source_language)
-    pipeline.set_pref("audio:target_language", target_language)
-    pipeline.describe()
-    pipeline.on(event_callback)
-
-    # Start processing audio
-    logger.info(f"Opening {filename}")
-    container = av.open(filename)
-    try:
-        logger.info("Start pushing audio into the pipeline")
-        for frame in container.decode(audio=0):
-            await pipeline.push(frame)
-    finally:
-        logger.info("Flushing the pipeline")
-        await pipeline.flush()
-
-    # Run diarization if enabled and we have topics
-    if enable_diarization and not only_transcript and audio_temp_path:
-        topics = topic_collector.get_topics()
-
-        if topics:
-            logger.info(f"Starting diarization with {len(topics)} topics")
-
-            try:
-                from reflector.processors import AudioDiarizationAutoProcessor
-
-                diarization_processor = AudioDiarizationAutoProcessor(
-                    name=diarization_backend
-                )
-
-                diarization_processor.set_pipeline(pipeline)
-
-                # For Modal backend, we need to upload the file to S3 first
-                if diarization_backend == "modal":
-                    from datetime import datetime
-
-                    from reflector.storage import get_transcripts_storage
-                    from reflector.utils.s3_temp_file import S3TemporaryFile
-
-                    storage = get_transcripts_storage()
-
-                    # Generate a unique filename in evaluation folder
-                    timestamp = datetime.utcnow().strftime("%Y%m%d_%H%M%S")
-                    audio_filename = f"evaluation/diarization_temp/{timestamp}_{uuid.uuid4().hex}.wav"
-
-                    # Use context manager for automatic cleanup
-                    async with S3TemporaryFile(storage, audio_filename) as s3_file:
-                        # Read and upload the audio file
-                        with open(audio_temp_path, "rb") as f:
-                            audio_data = f.read()
-
-                        audio_url = await s3_file.upload(audio_data)
-                        logger.info(f"Uploaded audio to S3: {audio_filename}")
-
-                        # Create diarization input with S3 URL
-                        diarization_input = AudioDiarizationInput(
-                            audio_url=audio_url, topics=topics
-                        )
-
-                        # Run diarization
-                        await diarization_processor.push(diarization_input)
-                        await diarization_processor.flush()
-
-                        logger.info("Diarization complete")
-                        # File will be automatically cleaned up when exiting the context
-                else:
-                    # For local backend, use local file path
-                    audio_url = audio_temp_path
-
-                    # Create diarization input
-                    diarization_input = AudioDiarizationInput(
-                        audio_url=audio_url, topics=topics
-                    )
-
-                    # Run diarization
-                    await diarization_processor.push(diarization_input)
-                    await diarization_processor.flush()
-
-                    logger.info("Diarization complete")
-
-            except ImportError as e:
-                logger.error(f"Failed to import diarization dependencies: {e}")
-                logger.error(
-                    "Install with: uv pip install pyannote.audio torch torchaudio"
-                )
-                logger.error(
-                    "And set HF_TOKEN environment variable for pyannote models"
-                )
-                raise SystemExit(1)
-            except Exception as e:
-                logger.error(f"Diarization failed: {e}")
-                raise SystemExit(1)
-        else:
-            logger.warning("Skipping diarization: no topics available")
-
-    # Clean up temp file
-    if audio_temp_path:
-        try:
-            Path(audio_temp_path).unlink()
-        except Exception as e:
-            logger.warning(f"Failed to clean up temp file {audio_temp_path}: {e}")
-
-    logger.info("All done!")
+    # result.ready() blocks even without await; it mutates result also
+    while not result.ready():
+        print(f"Status: {result.state}")
+        time.sleep(2)
 
 
 async def process_file_pipeline(
-    filename: str,
-    event_callback,
-    source_language="en",
-    target_language="en",
-    enable_diarization=True,
-    diarization_backend="modal",
+    transcript_id: TranscriptId,
 ):
     """Process audio/video file using the optimized file pipeline"""
+
+    # task_pipeline_file_process is a Celery task, need to use .delay() for async execution
+    result = task_pipeline_file_process.delay(transcript_id=transcript_id)
+
+    # Wait for the Celery task to complete
+    while not result.ready():
+        print(f"File pipeline status: {result.state}", file=sys.stderr)
+        time.sleep(2)
+
+    logger.info("File pipeline processing complete")
+
+
+async def process(
+    source_path: str,
+    source_language: str,
+    target_language: str,
+    pipeline: Literal["live", "file"],
+    output_path: str = None,
+):
+    from reflector.db import get_database
+
+    database = get_database()
+    # db connect is a part of ceremony
+    await database.connect()
+
     try:
-        from reflector.db import database
-        from reflector.db.transcripts import SourceKind, transcripts_controller
-        from reflector.pipelines.main_file_pipeline import PipelineMainFile
-
-        await database.connect()
-        try:
-            # Create a temporary transcript for processing
-            transcript = await transcripts_controller.add(
-                "",
-                source_kind=SourceKind.FILE,
-                source_language=source_language,
-                target_language=target_language,
-            )
-
-            # Process the file
-            pipeline = PipelineMainFile(transcript_id=transcript.id)
-            await pipeline.process(Path(filename))
-
-            logger.info("File pipeline processing complete")
-
-        finally:
-            await database.disconnect()
-    except ImportError as e:
-        logger.error(f"File pipeline not available: {e}")
-        logger.info("Falling back to stream pipeline")
-        # Fall back to stream pipeline
-        await process_audio_file(
-            filename,
-            event_callback,
-            only_transcript=False,
-            source_language=source_language,
-            target_language=target_language,
-            enable_diarization=enable_diarization,
-            diarization_backend=diarization_backend,
+        transcript_id = await prepare_entry(
+            source_path,
+            source_language,
+            target_language,
         )
 
+        pipeline_handlers = {
+            "live": process_live_pipeline,
+            "file": process_file_pipeline,
+        }
+
+        handler = pipeline_handlers.get(pipeline)
+        if not handler:
+            raise ValueError(f"Unknown pipeline type: {pipeline}")
+
+        await handler(transcript_id)
+
+        await extract_result_from_entry(transcript_id, output_path)
+    finally:
+        await database.disconnect()
+
 
 if __name__ == "__main__":
-    import argparse
-    import os
-
     parser = argparse.ArgumentParser(
-        description="Process audio files with optional speaker diarization"
+        description="Process audio files with speaker diarization"
     )
     parser.add_argument("source", help="Source file (mp3, wav, mp4...)")
     parser.add_argument(
-        "--stream",
-        action="store_true",
-        help="Use streaming pipeline (original frame-based processing)",
-    )
-    parser.add_argument(
-        "--only-transcript",
-        "-t",
-        action="store_true",
-        help="Only generate transcript without topics/summaries",
+        "--pipeline",
+        required=True,
+        choices=["live", "file"],
+        help="Pipeline type to use for processing (live: streaming/incremental, file: batch/parallel)",
     )
     parser.add_argument(
         "--source-language", default="en", help="Source language code (default: en)"
@@ -295,81 +207,14 @@ if __name__ == "__main__":
         "--target-language", default="en", help="Target language code (default: en)"
     )
     parser.add_argument("--output", "-o", help="Output file (output.jsonl)")
-    parser.add_argument(
-        "--enable-diarization",
-        "-d",
-        action="store_true",
-        help="Enable speaker diarization",
-    )
-    parser.add_argument(
-        "--diarization-backend",
-        default="pyannote",
-        choices=["pyannote", "modal"],
-        help="Diarization backend to use (default: pyannote)",
-    )
     args = parser.parse_args()
 
-    if "REDIS_HOST" not in os.environ:
-        os.environ["REDIS_HOST"] = "localhost"
-
-    output_fd = None
-    if args.output:
-        output_fd = open(args.output, "w")
-
-    async def event_callback(event: PipelineEvent):
-        processor = event.processor
-        data = event.data
-
-        # Ignore internal processors
-        if processor in (
-            "AudioChunkerProcessor",
-            "AudioMergeProcessor",
-            "AudioFileWriterProcessor",
-            "TopicCollectorProcessor",
-            "BroadcastProcessor",
-        ):
-            return
-
-        # If diarization is enabled, skip the original topic events from the pipeline
-        # The diarization processor will emit the same topics but with speaker info
-        if processor == "TranscriptTopicDetectorProcessor" and args.enable_diarization:
-            return
-
-        # Log all events
-        logger.info(f"Event: {processor} - {type(data).__name__}")
-
-        # Write to output
-        if output_fd:
-            output_fd.write(event.model_dump_json())
-            output_fd.write("\n")
-            output_fd.flush()
-
-    if args.stream:
-        # Use original streaming pipeline
-        asyncio.run(
-            process_audio_file(
-                args.source,
-                event_callback,
-                only_transcript=args.only_transcript,
-                source_language=args.source_language,
-                target_language=args.target_language,
-                enable_diarization=args.enable_diarization,
-                diarization_backend=args.diarization_backend,
-            )
+    asyncio.run(
+        process(
+            args.source,
+            args.source_language,
+            args.target_language,
+            args.pipeline,
+            args.output,
         )
-    else:
-        # Use optimized file pipeline (default)
-        asyncio.run(
-            process_file_pipeline(
-                args.source,
-                event_callback,
-                source_language=args.source_language,
-                target_language=args.target_language,
-                enable_diarization=args.enable_diarization,
-                diarization_backend=args.diarization_backend,
-            )
-        )
-
-    if output_fd:
-        output_fd.close()
-        logger.info(f"Output written to {args.output}")
+    )

server/reflector/tools/process_with_diarization.py

@@ -1,315 +0,0 @@
-"""
-@vibe-generated
-Process audio file with diarization support
-===========================================
-
-Extended version of process.py that includes speaker diarization.
-This tool processes audio files locally without requiring the full server infrastructure.
-"""
-
-import asyncio
-import tempfile
-import uuid
-from pathlib import Path
-from typing import List
-
-import av
-
-from reflector.logger import logger
-from reflector.processors import (
-    AudioChunkerProcessor,
-    AudioFileWriterProcessor,
-    AudioMergeProcessor,
-    AudioTranscriptAutoProcessor,
-    Pipeline,
-    PipelineEvent,
-    TranscriptFinalSummaryProcessor,
-    TranscriptFinalTitleProcessor,
-    TranscriptLinerProcessor,
-    TranscriptTopicDetectorProcessor,
-    TranscriptTranslatorAutoProcessor,
-)
-from reflector.processors.base import BroadcastProcessor, Processor
-from reflector.processors.types import (
-    AudioDiarizationInput,
-    TitleSummary,
-    TitleSummaryWithId,
-)
-
-
-class TopicCollectorProcessor(Processor):
-    """Collect topics for diarization"""
-
-    INPUT_TYPE = TitleSummary
-    OUTPUT_TYPE = TitleSummary
-
-    def __init__(self, **kwargs):
-        super().__init__(**kwargs)
-        self.topics: List[TitleSummaryWithId] = []
-        self._topic_id = 0
-
-    async def _push(self, data: TitleSummary):
-        # Convert to TitleSummaryWithId and collect
-        self._topic_id += 1
-        topic_with_id = TitleSummaryWithId(
-            id=str(self._topic_id),
-            title=data.title,
-            summary=data.summary,
-            timestamp=data.timestamp,
-            duration=data.duration,
-            transcript=data.transcript,
-        )
-        self.topics.append(topic_with_id)
-
-        # Pass through the original topic
-        await self.emit(data)
-
-    def get_topics(self) -> List[TitleSummaryWithId]:
-        return self.topics
-
-
-async def process_audio_file_with_diarization(
-    filename,
-    event_callback,
-    only_transcript=False,
-    source_language="en",
-    target_language="en",
-    enable_diarization=True,
-    diarization_backend="modal",
-):
-    # Create temp file for audio if diarization is enabled
-    audio_temp_path = None
-    if enable_diarization:
-        audio_temp_file = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
-        audio_temp_path = audio_temp_file.name
-        audio_temp_file.close()
-
-    # Create processor for collecting topics
-    topic_collector = TopicCollectorProcessor()
-
-    # Build pipeline for audio processing
-    processors = []
-
-    # Add audio file writer at the beginning if diarization is enabled
-    if enable_diarization:
-        processors.append(AudioFileWriterProcessor(audio_temp_path))
-
-    # Add the rest of the processors
-    processors += [
-        AudioChunkerProcessor(),
-        AudioMergeProcessor(),
-        AudioTranscriptAutoProcessor.as_threaded(),
-    ]
-
-    processors += [
-        TranscriptLinerProcessor(),
-        TranscriptTranslatorAutoProcessor.as_threaded(),
-    ]
-
-    if not only_transcript:
-        processors += [
-            TranscriptTopicDetectorProcessor.as_threaded(),
-            # Collect topics for diarization
-            topic_collector,
-            BroadcastProcessor(
-                processors=[
-                    TranscriptFinalTitleProcessor.as_threaded(),
-                    TranscriptFinalSummaryProcessor.as_threaded(),
-                ],
-            ),
-        ]
-
-    # Create main pipeline
-    pipeline = Pipeline(*processors)
-    pipeline.set_pref("audio:source_language", source_language)
-    pipeline.set_pref("audio:target_language", target_language)
-    pipeline.describe()
-    pipeline.on(event_callback)
-
-    # Start processing audio
-    logger.info(f"Opening {filename}")
-    container = av.open(filename)
-    try:
-        logger.info("Start pushing audio into the pipeline")
-        for frame in container.decode(audio=0):
-            await pipeline.push(frame)
-    finally:
-        logger.info("Flushing the pipeline")
-        await pipeline.flush()
-
-    # Run diarization if enabled and we have topics
-    if enable_diarization and not only_transcript and audio_temp_path:
-        topics = topic_collector.get_topics()
-
-        if topics:
-            logger.info(f"Starting diarization with {len(topics)} topics")
-
-            try:
-                from reflector.processors import AudioDiarizationAutoProcessor
-
-                diarization_processor = AudioDiarizationAutoProcessor(
-                    name=diarization_backend
-                )
-
-                diarization_processor.set_pipeline(pipeline)
-
-                # For Modal backend, we need to upload the file to S3 first
-                if diarization_backend == "modal":
-                    from datetime import datetime, timezone
-
-                    from reflector.storage import get_transcripts_storage
-                    from reflector.utils.s3_temp_file import S3TemporaryFile
-
-                    storage = get_transcripts_storage()
-
-                    # Generate a unique filename in evaluation folder
-                    timestamp = datetime.now(timezone.utc).strftime("%Y%m%d_%H%M%S")
-                    audio_filename = f"evaluation/diarization_temp/{timestamp}_{uuid.uuid4().hex}.wav"
-
-                    # Use context manager for automatic cleanup
-                    async with S3TemporaryFile(storage, audio_filename) as s3_file:
-                        # Read and upload the audio file
-                        with open(audio_temp_path, "rb") as f:
-                            audio_data = f.read()
-
-                        audio_url = await s3_file.upload(audio_data)
-                        logger.info(f"Uploaded audio to S3: {audio_filename}")
-
-                        # Create diarization input with S3 URL
-                        diarization_input = AudioDiarizationInput(
-                            audio_url=audio_url, topics=topics
-                        )
-
-                        # Run diarization
-                        await diarization_processor.push(diarization_input)
-                        await diarization_processor.flush()
-
-                        logger.info("Diarization complete")
-                        # File will be automatically cleaned up when exiting the context
-                else:
-                    # For local backend, use local file path
-                    audio_url = audio_temp_path
-
-                    # Create diarization input
-                    diarization_input = AudioDiarizationInput(
-                        audio_url=audio_url, topics=topics
-                    )
-
-                    # Run diarization
-                    await diarization_processor.push(diarization_input)
-                    await diarization_processor.flush()
-
-                    logger.info("Diarization complete")
-
-            except ImportError as e:
-                logger.error(f"Failed to import diarization dependencies: {e}")
-                logger.error(
-                    "Install with: uv pip install pyannote.audio torch torchaudio"
-                )
-                logger.error(
-                    "And set HF_TOKEN environment variable for pyannote models"
-                )
-                raise SystemExit(1)
-            except Exception as e:
-                logger.error(f"Diarization failed: {e}")
-                raise SystemExit(1)
-        else:
-            logger.warning("Skipping diarization: no topics available")
-
-    # Clean up temp file
-    if audio_temp_path:
-        try:
-            Path(audio_temp_path).unlink()
-        except Exception as e:
-            logger.warning(f"Failed to clean up temp file {audio_temp_path}: {e}")
-
-    logger.info("All done!")
-
-
-if __name__ == "__main__":
-    import argparse
-    import os
-
-    parser = argparse.ArgumentParser(
-        description="Process audio files with optional speaker diarization"
-    )
-    parser.add_argument("source", help="Source file (mp3, wav, mp4...)")
-    parser.add_argument(
-        "--only-transcript",
-        "-t",
-        action="store_true",
-        help="Only generate transcript without topics/summaries",
-    )
-    parser.add_argument(
-        "--source-language", default="en", help="Source language code (default: en)"
-    )
-    parser.add_argument(
-        "--target-language", default="en", help="Target language code (default: en)"
-    )
-    parser.add_argument("--output", "-o", help="Output file (output.jsonl)")
-    parser.add_argument(
-        "--enable-diarization",
-        "-d",
-        action="store_true",
-        help="Enable speaker diarization",
-    )
-    parser.add_argument(
-        "--diarization-backend",
-        default="modal",
-        choices=["modal"],
-        help="Diarization backend to use (default: modal)",
-    )
-    args = parser.parse_args()
-
-    # Set REDIS_HOST to localhost if not provided
-    if "REDIS_HOST" not in os.environ:
-        os.environ["REDIS_HOST"] = "localhost"
-        logger.info("REDIS_HOST not set, defaulting to localhost")
-
-    output_fd = None
-    if args.output:
-        output_fd = open(args.output, "w")
-
-    async def event_callback(event: PipelineEvent):
-        processor = event.processor
-        data = event.data
-
-        # Ignore internal processors
-        if processor in (
-            "AudioChunkerProcessor",
-            "AudioMergeProcessor",
-            "AudioFileWriterProcessor",
-            "TopicCollectorProcessor",
-            "BroadcastProcessor",
-        ):
-            return
-
-        # If diarization is enabled, skip the original topic events from the pipeline
-        # The diarization processor will emit the same topics but with speaker info
-        if processor == "TranscriptTopicDetectorProcessor" and args.enable_diarization:
-            return
-
-        # Log all events
-        logger.info(f"Event: {processor} - {type(data).__name__}")
-
-        # Write to output
-        if output_fd:
-            output_fd.write(event.model_dump_json())
-            output_fd.write("\n")
-            output_fd.flush()
-
-    asyncio.run(
-        process_audio_file_with_diarization(
-            args.source,
-            event_callback,
-            only_transcript=args.only_transcript,
-            source_language=args.source_language,
-            target_language=args.target_language,
-            enable_diarization=args.enable_diarization,
-            diarization_backend=args.diarization_backend,
-        )
-    )
-
-    if output_fd:
-        output_fd.close()
-        logger.info(f"Output written to {args.output}")

server/reflector/tools/runpipeline.py

@@ -53,7 +53,7 @@ async def run_single_processor(args):
     async def event_callback(event: PipelineEvent):
         processor = event.processor
         # ignore some processor
-        if processor in ("AudioChunkerProcessor", "AudioMergeProcessor"):
+        if processor in ("AudioChunkerAutoProcessor", "AudioMergeProcessor"):
             return
         print(f"Event: {event}")
         if output_fd:

server/reflector/tools/test_diarization.py

@@ -1,96 +0,0 @@
-#!/usr/bin/env python3
-"""
-@vibe-generated
-Test script for the diarization CLI tool
-=========================================
-
-This script helps test the diarization functionality with sample audio files.
-"""
-
-import asyncio
-import sys
-from pathlib import Path
-
-from reflector.logger import logger
-
-
-async def test_diarization(audio_file: str):
-    """Test the diarization functionality"""
-
-    # Import the processing function
-    from process_with_diarization import process_audio_file_with_diarization
-
-    # Collect events
-    events = []
-
-    async def event_callback(event):
-        events.append({"processor": event.processor, "data": event.data})
-        logger.info(f"Event from {event.processor}")
-
-    # Process the audio file
-    logger.info(f"Processing audio file: {audio_file}")
-
-    try:
-        await process_audio_file_with_diarization(
-            audio_file,
-            event_callback,
-            only_transcript=False,
-            source_language="en",
-            target_language="en",
-            enable_diarization=True,
-            diarization_backend="modal",
-        )
-
-        # Analyze results
-        logger.info(f"Processing complete. Received {len(events)} events")
-
-        # Look for diarization results
-        diarized_topics = []
-        for event in events:
-            if "TitleSummary" in event["processor"]:
-                # Check if words have speaker information
-                if hasattr(event["data"], "transcript") and event["data"].transcript:
-                    words = event["data"].transcript.words
-                    if words and hasattr(words[0], "speaker"):
-                        speakers = set(
-                            w.speaker for w in words if hasattr(w, "speaker")
-                        )
-                        logger.info(
-                            f"Found {len(speakers)} speakers in topic: {event['data'].title}"
-                        )
-                        diarized_topics.append(event["data"])
-
-        if diarized_topics:
-            logger.info(f"Successfully diarized {len(diarized_topics)} topics")
-
-            # Print sample output
-            sample_topic = diarized_topics[0]
-            logger.info("Sample diarized output:")
-            for i, word in enumerate(sample_topic.transcript.words[:10]):
-                logger.info(f"  Word {i}: '{word.text}' - Speaker {word.speaker}")
-        else:
-            logger.warning("No diarization results found in output")
-
-        return events
-
-    except Exception as e:
-        logger.error(f"Error during processing: {e}")
-        raise
-
-
-def main():
-    if len(sys.argv) < 2:
-        print("Usage: python test_diarization.py <audio_file>")
-        sys.exit(1)
-
-    audio_file = sys.argv[1]
-    if not Path(audio_file).exists():
-        print(f"Error: Audio file '{audio_file}' not found")
-        sys.exit(1)
-
-    # Run the test
-    asyncio.run(test_diarization(audio_file))
-
-
-if __name__ == "__main__":
-    main()

server/reflector/utils/string.py

@@ -0,0 +1,23 @@
+from typing import Annotated
+
+from pydantic import Field, TypeAdapter, constr
+
+NonEmptyStringBase = constr(min_length=1, strip_whitespace=False)
+NonEmptyString = Annotated[
+    NonEmptyStringBase,
+    Field(description="A non-empty string", min_length=1),
+]
+non_empty_string_adapter = TypeAdapter(NonEmptyString)
+
+
+def parse_non_empty_string(s: str, error: str | None = None) -> NonEmptyString:
+    try:
+        return non_empty_string_adapter.validate_python(s)
+    except Exception as e:
+        raise ValueError(f"{e}: {error}" if error else e) from e
+
+
+def try_parse_non_empty_string(s: str) -> NonEmptyString | None:
+    if not s:
+        return None
+    return parse_non_empty_string(s)

server/reflector/video_platforms/__init__.py

@@ -0,0 +1,18 @@
+# Video Platform Abstraction Layer
+"""
+This module provides an abstraction layer for different video conferencing platforms.
+It allows seamless switching between providers (Whereby, Daily.co, etc.) without
+changing the core application logic.
+"""
+
+from .base import VideoPlatformClient
+from .models import MeetingData, VideoPlatformConfig
+from .registry import get_platform_client, register_platform
+
+__all__ = [
+    "VideoPlatformClient",
+    "VideoPlatformConfig",
+    "MeetingData",
+    "get_platform_client",
+    "register_platform",
+]

server/reflector/video_platforms/base.py

@@ -0,0 +1,60 @@
+from abc import ABC, abstractmethod
+from datetime import datetime
+from typing import TYPE_CHECKING, Any, Dict, Optional
+
+from reflector.platform_types import Platform
+
+from .models import MeetingData, VideoPlatformConfig
+
+if TYPE_CHECKING:
+    from reflector.db.rooms import Room
+
+
+class VideoPlatformClient(ABC):
+    """Abstract base class for video platform integrations."""
+
+    PLATFORM_NAME: Platform
+
+    def __init__(self, config: VideoPlatformConfig):
+        self.config = config
+
+    @abstractmethod
+    async def create_meeting(
+        self, room_name_prefix: str, end_date: datetime, room: "Room"
+    ) -> MeetingData:
+        """Create a new meeting room."""
+        pass
+
+    @abstractmethod
+    async def get_room_sessions(self, room_name: str) -> Dict[str, Any]:
+        """Get session information for a room."""
+        pass
+
+    @abstractmethod
+    async def delete_room(self, room_name: str) -> bool:
+        """Delete a room. Returns True if successful."""
+        pass
+
+    @abstractmethod
+    async def upload_logo(self, room_name: str, logo_path: str) -> bool:
+        """Upload a logo to the room. Returns True if successful."""
+        pass
+
+    @abstractmethod
+    def verify_webhook_signature(
+        self, body: bytes, signature: str, timestamp: Optional[str] = None
+    ) -> bool:
+        """Verify webhook signature for security."""
+        pass
+
+    def format_recording_config(self, room: "Room") -> Dict[str, Any]:
+        """Format recording configuration for the platform.
+        Can be overridden by specific implementations."""
+        if room.recording_type == "cloud" and self.config.s3_bucket:
+            return {
+                "type": room.recording_type,
+                "bucket": self.config.s3_bucket,
+                "region": self.config.s3_region,
+                "trigger": room.recording_trigger,
+            }
+        return {"type": room.recording_type}