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/pull_request_template.md

@@ -1,19 +1,21 @@
-## ⚠️ Insert the PR TITLE replacing this text ⚠️
+<!--- Provide a general summary of your changes in the Title above -->
 
-⚠️ Describe your PR replacing this text. Post screenshots or videos whenever possible. ⚠️
+## Description
+<!--- Describe your changes in detail -->
 
-### Checklist
+## Related Issue
+<!--- This project only accepts pull requests related to open issues -->
+<!--- If suggesting a new feature or change, please discuss it in an issue first -->
+<!--- If fixing a bug, there should be an issue describing it with steps to reproduce -->
+<!--- Please link to the issue here: -->
 
- - [ ] My branch is updated with main (mandatory)
- - [ ] I wrote unit tests for this (if applies)
- - [ ] I have included migrations and tested them locally (if applies)
- - [ ] I have manually tested this feature locally
+## Motivation and Context
+<!--- Why is this change required? What problem does it solve? -->
+<!--- If it fixes an open issue, please link to the issue here. -->
 
-> IMPORTANT: Remember that you are responsible for merging this PR after it's been reviewed, and once deployed
-> you should perform manual testing to make sure everything went smoothly.
-
-### Urgency
-
- - [ ] Urgent (deploy ASAP)
- - [ ] Non-urgent (deploying in next release is ok)
+## How Has This Been Tested?
+<!--- Please describe in detail how you tested your changes. -->
+<!--- Include details of your testing environment, and the tests you ran to -->
+<!--- see how your change affects other areas of the code, etc. -->
 
+## Screenshots (if appropriate):

.github/workflows/conventional_commit_pr_title.yml

@@ -0,0 +1,21 @@
+name: "Lint PR"
+
+on:
+  pull_request_target:
+    types:
+      - opened
+      - edited
+      - synchronize
+      - reopened
+
+permissions:
+  pull-requests: read
+
+jobs:
+  main:
+    name: Validate PR title
+    runs-on: ubuntu-latest
+    steps:
+      - uses: amannn/action-semantic-pull-request@v5
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}

.github/workflows/db_migrations.yml

@@ -2,6 +2,8 @@ name: Test Database Migrations
 
 on:
   push:
+    branches:
+      - main
     paths:
       - "server/migrations/**"
       - "server/reflector/db/**"
@@ -17,39 +19,63 @@ on:
 jobs:
   test-migrations:
     runs-on: ubuntu-latest
+    concurrency:
+      group: db-ubuntu-latest-${{ github.ref }}
+      cancel-in-progress: true
+    services:
+      postgres:
+        image: postgres:17
+        env:
+          POSTGRES_USER: reflector
+          POSTGRES_PASSWORD: reflector
+          POSTGRES_DB: reflector
+        ports:
+          - 5432:5432
+        options: >-
+          --health-cmd pg_isready -h 127.0.0.1 -p 5432
+          --health-interval 10s
+          --health-timeout 5s
+          --health-retries 5
+
+    env:
+      DATABASE_URL: postgresql://reflector:reflector@localhost:5432/reflector
 
     steps:
       - uses: actions/checkout@v4
 
-      - name: Install poetry
-        run: pipx install poetry
+      - name: Install PostgreSQL client
+        run: sudo apt-get update && sudo apt-get install -y postgresql-client | cat
 
-      - name: Set up Python 3.x
-        uses: actions/setup-python@v4
-        with:
-          python-version: "3.11"
-          cache: "poetry"
-          cache-dependency-path: "server/poetry.lock"
-
-      - name: Install requirements
-        working-directory: ./server
+      - name: Wait for Postgres
         run: |
-          poetry install --no-root
+          for i in {1..30}; do
+            if pg_isready -h localhost -p 5432; then
+              echo "Postgres is ready"
+              break
+            fi
+            echo "Waiting for Postgres... ($i)" && sleep 1
+          done
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@v3
+        with:
+          enable-cache: true
+          working-directory: server
 
       - name: Test migrations from scratch
-        working-directory: ./server
+        working-directory: server
         run: |
           echo "Testing migrations from clean database..."
-          poetry run alembic upgrade head
+          uv run alembic upgrade head
           echo "✅ Fresh migration successful"
 
       - name: Test migration rollback and re-apply
-        working-directory: ./server
+        working-directory: server
         run: |
           echo "Testing rollback to base..."
-          poetry run alembic downgrade base
+          uv run alembic downgrade base
           echo "✅ Rollback successful"
 
           echo "Testing re-apply of all migrations..."
-          poetry run alembic upgrade head
+          uv run alembic upgrade head
           echo "✅ Re-apply successful"

.github/workflows/deploy.yml

@@ -8,18 +8,30 @@ env:
   ECR_REPOSITORY: reflector
 
 jobs:
-  deploy:
-    runs-on: ubuntu-latest
+  build:
+    strategy:
+      matrix:
+        include:
+          - platform: linux/amd64
+            runner: linux-amd64
+            arch: amd64
+          - platform: linux/arm64
+            runner: linux-arm64
+            arch: arm64
+
+    runs-on: ${{ matrix.runner }}
 
     permissions:
-      deployments: write
       contents: read
 
+    outputs:
+      registry: ${{ steps.login-ecr.outputs.registry }}
+
     steps:
-      - uses: actions/checkout@v3
+      - uses: actions/checkout@v4
 
       - name: Configure AWS credentials
-        uses: aws-actions/configure-aws-credentials@0e613a0980cbf65ed5b322eb7a1e075d28913a83
+        uses: aws-actions/configure-aws-credentials@v4
         with:
           aws-access-key-id: ${{ secrets.AWS_ACCESS_KEY_ID }}
           aws-secret-access-key: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
@@ -27,21 +39,52 @@ jobs:
 
       - name: Login to Amazon ECR
         id: login-ecr
-        uses: aws-actions/amazon-ecr-login@62f4f872db3836360b72999f4b87f1ff13310f3a
-
-      - name: Set up QEMU
-        uses: docker/setup-qemu-action@v2
+        uses: aws-actions/amazon-ecr-login@v2
 
       - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v2
+        uses: docker/setup-buildx-action@v3
 
-      - name: Build and push
-        id: docker_build
-        uses: docker/build-push-action@v4
+      - name: Build and push ${{ matrix.arch }}
+        uses: docker/build-push-action@v5
         with:
           context: server
-          platforms: linux/amd64,linux/arm64
+          platforms: ${{ matrix.platform }}
           push: true
-          tags: ${{ steps.login-ecr.outputs.registry }}/${{ env.ECR_REPOSITORY }}:latest
-          cache-from: type=gha
-          cache-to: type=gha,mode=max
+          tags: ${{ steps.login-ecr.outputs.registry }}/${{ env.ECR_REPOSITORY }}:latest-${{ matrix.arch }}
+          cache-from: type=gha,scope=${{ matrix.arch }}
+          cache-to: type=gha,mode=max,scope=${{ matrix.arch }}
+          github-token: ${{ secrets.GHA_CACHE_TOKEN }}
+          provenance: false
+
+  create-manifest:
+    runs-on: ubuntu-latest
+    needs: [build]
+
+    permissions:
+      deployments: write
+      contents: read
+
+    steps:
+      - name: Configure AWS credentials
+        uses: aws-actions/configure-aws-credentials@v4
+        with:
+          aws-access-key-id: ${{ secrets.AWS_ACCESS_KEY_ID }}
+          aws-secret-access-key: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
+          aws-region: ${{ env.AWS_REGION }}
+
+      - name: Login to Amazon ECR
+        uses: aws-actions/amazon-ecr-login@v2
+
+      - name: Create and push multi-arch manifest
+        run: |
+          # Get the registry URL (since we can't easily access job outputs in matrix)
+          ECR_REGISTRY=$(aws ecr describe-registry --query 'registryId' --output text).dkr.ecr.${{ env.AWS_REGION }}.amazonaws.com
+
+          docker manifest create \
+            $ECR_REGISTRY/${{ env.ECR_REPOSITORY }}:latest \
+            $ECR_REGISTRY/${{ env.ECR_REPOSITORY }}:latest-amd64 \
+            $ECR_REGISTRY/${{ env.ECR_REPOSITORY }}:latest-arm64
+
+          docker manifest push $ECR_REGISTRY/${{ env.ECR_REPOSITORY }}:latest
+
+          echo "✅ Multi-arch manifest pushed: $ECR_REGISTRY/${{ env.ECR_REPOSITORY }}:latest"

.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/pre-commit.yml

@@ -0,0 +1,24 @@
+name: pre-commit
+
+on:
+  pull_request:
+  push:
+    branches: [main]
+
+jobs:
+  pre-commit:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v5
+      - uses: actions/setup-python@v5
+      - uses: pnpm/action-setup@v4
+        with:
+          version: 10
+      - uses: actions/setup-node@v4
+        with:
+          node-version: 22
+          cache: "pnpm"
+          cache-dependency-path: "www/pnpm-lock.yaml"
+      - name: Install dependencies
+        run: cd www && pnpm install --frozen-lockfile
+      - uses: pre-commit/action@v3.0.1

.github/workflows/release-please.yml

@@ -0,0 +1,19 @@
+on:
+  push:
+    branches:
+      - main
+
+permissions:
+  contents: write
+  pull-requests: write
+
+name: release-please
+
+jobs:
+  release-please:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: googleapis/release-please-action@v4
+        with:
+          token: ${{ secrets.MY_RELEASE_PLEASE_TOKEN }}
+          release-type: simple

.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

@@ -5,77 +5,66 @@ 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
         ports:
           - 6379:6379
     steps:
-      - uses: actions/checkout@v3
-      - name: Install poetry
-        run: pipx install poetry
-      - name: Set up Python 3.x
-        uses: actions/setup-python@v4
+      - uses: actions/checkout@v4
+      - name: Install uv
+        uses: astral-sh/setup-uv@v6
         with:
-          python-version: "3.11"
-          cache: "poetry"
-          cache-dependency-path: "server/poetry.lock"
-      - name: Install requirements
-        run: |
-          cd server
-          poetry install --no-root
+          enable-cache: true
+          working-directory: server
       - name: Tests
         run: |
           cd server
-          poetry run python -m pytest -v tests
+          uv run -m pytest -v tests
 
-  formatting:
-    runs-on: ubuntu-latest
+  docker-amd64:
+    runs-on: linux-amd64
+    concurrency:
+      group: docker-amd64-${{ github.ref }}
+      cancel-in-progress: true
     steps:
-      - uses: actions/checkout@v3
-      - name: Set up Python 3.x
-        uses: actions/setup-python@v4
-        with:
-          python-version: 3.11
-      - name: Validate formatting
-        run: |
-          pip install black
-          cd server
-          black --check reflector tests
-
-  linting:
-    runs-on: ubuntu-latest
-    steps:
-      - uses: actions/checkout@v3
-      - name: Set up Python 3.x
-        uses: actions/setup-python@v4
-        with:
-          python-version: 3.11
-      - name: Validate formatting
-        run: |
-          pip install ruff
-          cd server
-          ruff check reflector tests
-
-  docker:
-    runs-on: ubuntu-latest
-    steps:
-      - uses: actions/checkout@v3
-      - name: Set up QEMU
-        uses: docker/setup-qemu-action@v2
+      - uses: actions/checkout@v4
       - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v2
-      - name: Build and push
-        id: docker_build
-        uses: docker/build-push-action@v4
+        uses: docker/setup-buildx-action@v3
+      - name: Build AMD64
+        uses: docker/build-push-action@v6
         with:
           context: server
-          platforms: linux/amd64,linux/arm64
-          cache-from: type=gha
-          cache-to: type=gha,mode=max
+          platforms: linux/amd64
+          cache-from: type=gha,scope=amd64
+          cache-to: type=gha,mode=max,scope=amd64
+          github-token: ${{ secrets.GHA_CACHE_TOKEN }}
+
+  docker-arm64:
+    runs-on: linux-arm64
+    concurrency:
+      group: docker-arm64-${{ github.ref }}
+      cancel-in-progress: true
+    steps:
+      - uses: actions/checkout@v4
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+      - name: Build ARM64
+        uses: docker/build-push-action@v6
+        with:
+          context: server
+          platforms: linux/arm64
+          cache-from: type=gha,scope=arm64
+          cache-to: type=gha,mode=max,scope=arm64
+          github-token: ${{ secrets.GHA_CACHE_TOKEN }}

.gitignore

@@ -9,4 +9,12 @@ dump.rdb
 ngrok.log
 .claude/settings.local.json
 restart-dev.sh
-*.log
+*.log
+data/
+www/REFACTOR.md
+www/reload-frontend
+server/test.sqlite
+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

@@ -3,10 +3,10 @@
 repos:
   - repo: local
     hooks:
-      - id: yarn-format
-        name: run yarn format
+      - id: format
+        name: run format
         language: system
-        entry: bash -c 'cd www && yarn format'
+        entry: bash -c 'cd www && pnpm format'
         pass_filenames: false
         files: ^www/
 
@@ -15,25 +15,20 @@ repos:
     hooks:
       - id: debug-statements
       - id: trailing-whitespace
-        exclude: ^server/trials
       - id: detect-private-key
 
-  - repo: https://github.com/psf/black
-    rev: 24.1.1
-    hooks:
-      - id: black
-        files: ^server/(reflector|tests)/
-
-  - repo: https://github.com/pycqa/isort
-    rev: 5.12.0
-    hooks:
-      - id: isort
-        name: isort (python)
-        files: ^server/(gpu|evaluate|reflector)/
-        args: [ "--profile", "black", "--filter-files" ]
-
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.6.5
+    rev: v0.8.2
     hooks:
       - id: ruff
-        files: ^server/(reflector|tests)/
+        args:
+          - --fix
+          # Uses select rules from server/pyproject.toml
+        files: ^server/
+      - id: ruff-format
+        files: ^server/
+
+  - repo: https://github.com/gitleaks/gitleaks
+    rev: v8.28.0
+    hooks:
+      - id: gitleaks

.python-version

@@ -1 +0,0 @@
-3.11.6

CHANGELOG.md

@@ -0,0 +1,275 @@
+# 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)
+
+
+### Bug Fixes
+
+* webvtt db null expectation mismatch ([#556](https://github.com/Monadical-SAS/reflector/issues/556)) ([e67ad1a](https://github.com/Monadical-SAS/reflector/commit/e67ad1a4a2054467bfeb1e0258fbac5868aaaf21))
+
+## [0.7.0](https://github.com/Monadical-SAS/reflector/compare/v0.6.1...v0.7.0) (2025-08-21)
+
+
+### Features
+
+* delete recording with transcript ([#547](https://github.com/Monadical-SAS/reflector/issues/547)) ([99cc984](https://github.com/Monadical-SAS/reflector/commit/99cc9840b3f5de01e0adfbfae93234042d706d13))
+* pipeline improvement with file processing, parakeet, silero-vad ([#540](https://github.com/Monadical-SAS/reflector/issues/540)) ([bcc29c9](https://github.com/Monadical-SAS/reflector/commit/bcc29c9e0050ae215f89d460e9d645aaf6a5e486))
+* postgresql migration and removal of sqlite in pytest ([#546](https://github.com/Monadical-SAS/reflector/issues/546)) ([cd1990f](https://github.com/Monadical-SAS/reflector/commit/cd1990f8f0fe1503ef5069512f33777a73a93d7f))
+* search backend ([#537](https://github.com/Monadical-SAS/reflector/issues/537)) ([5f9b892](https://github.com/Monadical-SAS/reflector/commit/5f9b89260c9ef7f3c921319719467df22830453f))
+* search frontend  ([#551](https://github.com/Monadical-SAS/reflector/issues/551)) ([3657242](https://github.com/Monadical-SAS/reflector/commit/365724271ca6e615e3425125a69ae2b46ce39285))
+
+
+### Bug Fixes
+
+* evaluation cli event wrap ([#536](https://github.com/Monadical-SAS/reflector/issues/536)) ([941c3db](https://github.com/Monadical-SAS/reflector/commit/941c3db0bdacc7b61fea412f3746cc5a7cb67836))
+* use structlog not logging ([#550](https://github.com/Monadical-SAS/reflector/issues/550)) ([27e2f81](https://github.com/Monadical-SAS/reflector/commit/27e2f81fda5232e53edc729d3e99c5ef03adbfe9))
+
+## [0.6.1](https://github.com/Monadical-SAS/reflector/compare/v0.6.0...v0.6.1) (2025-08-06)
+
+
+### Bug Fixes
+
+* delayed waveform loading ([#538](https://github.com/Monadical-SAS/reflector/issues/538)) ([ef64146](https://github.com/Monadical-SAS/reflector/commit/ef64146325d03f64dd9a1fe40234fb3e7e957ae2))
+
+## [0.6.0](https://github.com/Monadical-SAS/reflector/compare/v0.5.0...v0.6.0) (2025-08-05)
+
+
+### ⚠ BREAKING CHANGES
+
+* Configuration keys have changed. Update your .env file:
+    - TRANSCRIPT_MODAL_API_KEY → TRANSCRIPT_API_KEY
+    - LLM_MODAL_API_KEY → (removed, use TRANSCRIPT_API_KEY)
+    - Add DIARIZATION_API_KEY and TRANSLATE_API_KEY if using those services
+
+### Features
+
+* implement service-specific Modal API keys with auto processor pattern ([#528](https://github.com/Monadical-SAS/reflector/issues/528)) ([650befb](https://github.com/Monadical-SAS/reflector/commit/650befb291c47a1f49e94a01ab37d8fdfcd2b65d))
+* use llamaindex everywhere ([#525](https://github.com/Monadical-SAS/reflector/issues/525)) ([3141d17](https://github.com/Monadical-SAS/reflector/commit/3141d172bc4d3b3d533370c8e6e351ea762169bf))
+
+
+### Miscellaneous Chores
+
+* **main:** release 0.6.0 ([ecdbf00](https://github.com/Monadical-SAS/reflector/commit/ecdbf003ea2476c3e95fd231adaeb852f2943df0))
+
+## [0.5.0](https://github.com/Monadical-SAS/reflector/compare/v0.4.0...v0.5.0) (2025-07-31)
+
+
+### Features
+
+* new summary using phi-4 and llama-index ([#519](https://github.com/Monadical-SAS/reflector/issues/519)) ([1bf9ce0](https://github.com/Monadical-SAS/reflector/commit/1bf9ce07c12f87f89e68a1dbb3b2c96c5ee62466))
+
+
+### Bug Fixes
+
+* remove unused settings and utils files ([#522](https://github.com/Monadical-SAS/reflector/issues/522)) ([2af4790](https://github.com/Monadical-SAS/reflector/commit/2af4790e4be9e588f282fbc1bb171c88a03d6479))
+
+## [0.4.0](https://github.com/Monadical-SAS/reflector/compare/v0.3.2...v0.4.0) (2025-07-25)
+
+
+### Features
+
+* Diarization cli ([#509](https://github.com/Monadical-SAS/reflector/issues/509)) ([ffc8003](https://github.com/Monadical-SAS/reflector/commit/ffc8003e6dad236930a27d0fe3e2f2adfb793890))
+
+
+### Bug Fixes
+
+* remove faulty import Meeting ([#512](https://github.com/Monadical-SAS/reflector/issues/512)) ([0e68c79](https://github.com/Monadical-SAS/reflector/commit/0e68c798434e1b481f9482cc3a4702ea00365df4))
+* room concurrency (theoretically) ([#511](https://github.com/Monadical-SAS/reflector/issues/511)) ([7bb3676](https://github.com/Monadical-SAS/reflector/commit/7bb367653afeb2778cff697a0eb217abf0b81b84))
+
+## [0.3.2](https://github.com/Monadical-SAS/reflector/compare/v0.3.1...v0.3.2) (2025-07-22)
+
+
+### Bug Fixes
+
+* match font size for the filter sidebar ([#507](https://github.com/Monadical-SAS/reflector/issues/507)) ([4b8ba5d](https://github.com/Monadical-SAS/reflector/commit/4b8ba5db1733557e27b098ad3d1cdecadf97ae52))
+* whereby consent not displaying ([#505](https://github.com/Monadical-SAS/reflector/issues/505)) ([1120552](https://github.com/Monadical-SAS/reflector/commit/1120552c2c83d084d3a39272ad49b6aeda1af98f))
+
+## [0.3.1](https://github.com/Monadical-SAS/reflector/compare/v0.3.0...v0.3.1) (2025-07-22)
+
+
+### Bug Fixes
+
+* remove fief out of the source code ([#502](https://github.com/Monadical-SAS/reflector/issues/502)) ([890dd15](https://github.com/Monadical-SAS/reflector/commit/890dd15ba5a2be10dbb841e9aeb75d377885f4af))
+* remove primary color for room action menu ([#504](https://github.com/Monadical-SAS/reflector/issues/504)) ([2e33f89](https://github.com/Monadical-SAS/reflector/commit/2e33f89c0f9e5fbaafa80e8d2ae9788450ea2f31))
+
+## [0.3.0](https://github.com/Monadical-SAS/reflector/compare/v0.2.1...v0.3.0) (2025-07-21)
+
+
+### Features
+
+* migrate from chakra 2 to chakra 3 ([#500](https://github.com/Monadical-SAS/reflector/issues/500)) ([a858464](https://github.com/Monadical-SAS/reflector/commit/a858464c7a80e5497acf801d933bf04092f8b526))
+
+## [0.2.1](https://github.com/Monadical-SAS/reflector/compare/v0.2.0...v0.2.1) (2025-07-18)
+
+
+### Bug Fixes
+
+* separate browsing page into different components, limit to 10 by default ([#498](https://github.com/Monadical-SAS/reflector/issues/498)) ([c752da6](https://github.com/Monadical-SAS/reflector/commit/c752da6b97c96318aff079a5b2a6eceadfbfcad1))
+
+## [0.2.0](https://github.com/Monadical-SAS/reflector/compare/0.1.1...v0.2.0) (2025-07-17)
+
+
+### Features
+
+* improve transcript listing with room_id ([#496](https://github.com/Monadical-SAS/reflector/issues/496)) ([d2b5de5](https://github.com/Monadical-SAS/reflector/commit/d2b5de543fc0617fc220caa6a8a290e4040cb10b))
+
+
+### Bug Fixes
+
+* don't attempt to load waveform/mp3 if audio was deleted ([#495](https://github.com/Monadical-SAS/reflector/issues/495)) ([f4578a7](https://github.com/Monadical-SAS/reflector/commit/f4578a743fd0f20312fbd242fa9cccdfaeb20a9e))
+
+## [0.1.1](https://github.com/Monadical-SAS/reflector/compare/0.1.0...v0.1.1) (2025-07-17)
+
+
+### Bug Fixes
+
+* postgres database not connecting in worker ([#492](https://github.com/Monadical-SAS/reflector/issues/492)) ([123d09f](https://github.com/Monadical-SAS/reflector/commit/123d09fdacef7f5a84541cf01732d4f5b6b9d2d0))
+* process meetings with utc ([#493](https://github.com/Monadical-SAS/reflector/issues/493)) ([f3c85e1](https://github.com/Monadical-SAS/reflector/commit/f3c85e1eb97cd893840125ed056dcb290fccb612))
+* punkt -> punkt_tab + pre-download nltk packages to prevent runtime not working ([#489](https://github.com/Monadical-SAS/reflector/issues/489)) ([c22487b](https://github.com/Monadical-SAS/reflector/commit/c22487b41f311a3fdba2eac04c7637bd396cccee))
+* rename averaged_perceptron_tagger to averaged_perceptron_tagger_eng ([#491](https://github.com/Monadical-SAS/reflector/issues/491)) ([a7b7846](https://github.com/Monadical-SAS/reflector/commit/a7b78462419b3af81c6dbf1ddfccb3d532f660a3))

CLAUDE.md

@@ -0,0 +1,179 @@
+# CLAUDE.md
+
+This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
+
+## Project Overview
+
+Reflector is an AI-powered audio transcription and meeting analysis platform with real-time processing capabilities. The system consists of:
+
+- **Frontend**: Next.js 14 React application (`www/`) with Chakra UI, real-time WebSocket integration
+- **Backend**: Python FastAPI server (`server/`) with async database operations and background processing
+- **Processing**: GPU-accelerated ML pipeline for transcription, diarization, summarization via Modal.com
+- **Infrastructure**: Redis, PostgreSQL/SQLite, Celery workers, WebRTC streaming
+
+## Development Commands
+
+### Backend (Python) - `cd server/`
+
+**Setup and Dependencies:**
+```bash
+# Install dependencies
+uv sync
+
+# Database migrations (first run or schema changes)
+uv run alembic upgrade head
+
+# Start services
+docker compose up -d redis
+```
+
+**Development:**
+```bash
+# Start FastAPI server
+uv run -m reflector.app --reload
+
+# Start Celery worker for background tasks
+uv run celery -A reflector.worker.app worker --loglevel=info
+
+# Start Celery beat scheduler (optional, for cron jobs)
+uv run celery -A reflector.worker.app beat
+```
+
+**Testing:**
+```bash
+# Run all tests with coverage
+uv run pytest
+
+# Run specific test file
+uv run pytest tests/test_transcripts.py
+
+# Run tests with verbose output
+uv run pytest -v
+```
+
+**Process Audio Files:**
+```bash
+# Process local audio file manually
+uv run python -m reflector.tools.process path/to/audio.wav
+```
+
+### Frontend (Next.js) - `cd www/`
+
+**Setup:**
+```bash
+# Install dependencies
+pnpm install
+
+# Copy configuration templates
+cp .env_template .env
+```
+
+**Development:**
+```bash
+# Start development server
+pnpm dev
+
+# Generate TypeScript API client from OpenAPI spec
+pnpm openapi
+
+# Lint code
+pnpm lint
+
+# Format code
+pnpm format
+
+# Build for production
+pnpm build
+```
+
+### Docker Compose (Full Stack)
+
+```bash
+# Start all services
+docker compose up -d
+
+# Start specific services
+docker compose up -d redis server worker
+```
+
+## Architecture Overview
+
+### Backend Processing Pipeline
+
+The audio processing follows a modular pipeline architecture:
+
+1. **Audio Input**: WebRTC streaming, file upload, or cloud recording ingestion
+2. **Chunking**: Audio split into processable segments (`AudioChunkerProcessor`)
+3. **Transcription**: Whisper or Modal.com GPU processing (`AudioTranscriptAutoProcessor`)
+4. **Diarization**: Speaker identification (`AudioDiarizationAutoProcessor`)
+5. **Text Processing**: Formatting, translation, topic detection
+6. **Summarization**: AI-powered summaries and title generation
+7. **Storage**: Database persistence with optional S3 backend
+
+### Database Models
+
+Core entities:
+- `transcript`: Main table with processing results, summaries, topics, participants
+- `meeting`: Live meeting sessions with consent management
+- `room`: Virtual meeting spaces with configuration
+- `recording`: Audio/video file metadata and processing status
+
+### API Structure
+
+All endpoints prefixed `/v1/`:
+- `transcripts/` - CRUD operations for transcripts
+- `transcripts_audio/` - Audio streaming and download
+- `transcripts_webrtc/` - Real-time WebRTC endpoints
+- `transcripts_websocket/` - WebSocket for live updates
+- `meetings/` - Meeting lifecycle management
+- `rooms/` - Virtual room management
+
+### Frontend Architecture
+
+- **App Router**: Next.js 14 with route groups for organization
+- **State**: React Context pattern, no Redux
+- **Real-time**: WebSocket integration for live transcription updates
+- **Auth**: NextAuth.js with Authentik OAuth/OIDC provider
+- **UI**: Chakra UI components with Tailwind CSS utilities
+
+## Key Configuration
+
+### Environment Variables
+
+**Backend** (`server/.env`):
+- `DATABASE_URL` - Database connection string
+- `REDIS_URL` - Redis broker for Celery
+- `TRANSCRIPT_BACKEND=modal` + `TRANSCRIPT_MODAL_API_KEY` - Modal.com transcription
+- `DIARIZATION_BACKEND=modal` + `DIARIZATION_MODAL_API_KEY` - Modal.com diarization
+- `TRANSLATION_BACKEND=modal` + `TRANSLATION_MODAL_API_KEY` - Modal.com translation
+- `WHEREBY_API_KEY` - Video platform integration
+- `REFLECTOR_AUTH_BACKEND` - Authentication method (none, jwt)
+
+**Frontend** (`www/.env`):
+- `NEXTAUTH_URL`, `NEXTAUTH_SECRET` - Authentication configuration
+- `REFLECTOR_API_URL` - Backend API endpoint
+- `REFLECTOR_DOMAIN_CONFIG` - Feature flags and domain settings
+
+## Testing Strategy
+
+- **Backend**: pytest with async support, HTTP client mocking, audio processing tests
+- **Frontend**: No current test suite - opportunities for Jest/React Testing Library
+- **Coverage**: Backend maintains test coverage reports in `htmlcov/`
+
+## GPU Processing
+
+Modal.com integration for scalable ML processing:
+- Deploy changes: `modal run server/gpu/path/to/model.py`
+- Requires Modal account with `REFLECTOR_GPU_APIKEY` secret
+- Fallback to local processing when Modal unavailable
+
+## Common Issues
+
+- **Permissions**: Browser microphone access required in System Preferences
+- **Audio Routing**: Use BlackHole (Mac) for merging multiple audio sources
+- **WebRTC**: Ensure proper CORS configuration for cross-origin streaming
+- **Database**: Run `uv run alembic upgrade head` after pulling schema changes
+
+## Pipeline/worker related info
+
+If you need to do any worker/pipeline related work, search for "Pipeline" classes and their "create" or "build" methods to find the main processor sequence. Look for task orchestration patterns (like "chord", "group", or "chain") to identify the post-processing flow with parallel execution chains. This will give you abstract vision on how processing pipeling is organized.

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.

LICENSE

@@ -0,0 +1,9 @@
+MIT License
+
+Copyright (c) 2025 Monadical SAS
+
+Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

README.md

@@ -1,48 +1,69 @@
+<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/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>
+</div>
+<table>
+  <tr>
+    <td>
+      <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/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/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 https://github.com/fief-dev for authentication, and Vercel for deployment and configuration of the front-end.
+It also uses authentik for authentication if activated.
 
-## Table of Contents
+## Contribution Guidelines
 
-- [Reflector](#reflector)
-  - [Table of Contents](#table-of-contents)
-  - [Miscellaneous](#miscellaneous)
-    - [Contribution Guidelines](#contribution-guidelines)
-    - [How to Install Blackhole (Mac Only)](#how-to-install-blackhole-mac-only)
-  - [Front-End](#front-end)
-    - [Installation](#installation)
-    - [Run the Application](#run-the-application)
-    - [OpenAPI Code Generation](#openapi-code-generation)
-  - [Back-End](#back-end)
-    - [Installation](#installation-1)
-    - [Start the API/Backend](#start-the-apibackend)
-    - [Redis (Mac)](#redis-mac)
-    - [Redis (Windows)](#redis-windows)
-  - [Update the database schema (run on first install, and after each pull containing a migration)](#update-the-database-schema-run-on-first-install-and-after-each-pull-containing-a-migration)
-  - [Main Server](#main-server)
-    - [Crontab (optional)](#crontab-optional)
-      - [Using docker](#using-docker)
-    - [Using local GPT4All](#using-local-gpt4all)
-    - [Using local files](#using-local-files)
-  - [AI Models](#ai-models)
+All new contributions should be made in a separate branch, and goes through a Pull Request.
+[Conventional commits](https://www.conventionalcommits.org/en/v1.0.0/) must be used for the PR title and commits.
 
-## Miscellaneous
+## Usage
 
-### Contribution Guidelines
-
-All new contributions should be made in a separate branch. Before any code is merged into `main`, it requires a code review.
-
-### Usage instructions
-
-To record both your voice and the meeting you're taking part in, you need :
+To record both your voice and the meeting you're taking part in, you need:
 
 - For an in-person meeting, make sure your microphone is in range of all participants.
 - If using several microphones, make sure to merge the audio feeds into one with an external tool.
@@ -66,156 +87,121 @@ Note: We currently do not have instructions for Windows users.
 - Then goto `System Preferences -> Sound` and choose the devices created from the Output and Input tabs.
 - The input from your local microphone, the browser run meeting should be aggregated into one virtual stream to listen to and the output should be fed back to your specified output devices if everything is configured properly.
 
-## Front-End
+## Installation
+
+*Note: we're working toward better installation, theses instructions are not accurate for now*
+
+### Frontend
 
 Start with `cd www`.
 
-### Installation
-
-To install the application, run:
+**Installation**
 
 ```bash
-yarn install
-cp .env_template .env
-cp config-template.ts config.ts
+pnpm install
+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 the Application
-
-To run the application in development mode, run:
+**Run in development mode**
 
 ```bash
-yarn dev
+pnpm dev
 ```
 
 Then (after completing server setup and starting it) open [http://localhost:3000](http://localhost:3000) to view it in the browser.
 
-### OpenAPI Code Generation
+**OpenAPI Code Generation**
 
 To generate the TypeScript files from the openapi.json file, make sure the python server is running, then run:
 
 ```bash
-yarn openapi
+pnpm openapi
 ```
 
-## Back-End
+### Backend
 
 Start with `cd server`.
 
-### Quick-run instructions (only if you installed everything already)
-
-```bash
-redis-server # Mac
-docker compose up -d redis # Windows
-poetry run celery -A reflector.worker.app worker --loglevel=info
-poetry run python -m reflector.app
-```
-
-### Installation
-
-Download [Python 3.11 from the official website](https://www.python.org/downloads/) and ensure you have version 3.11 by running `python --version`.
-
-Run:
-
-```bash
-python --version # It should say 3.11
-pip install poetry
-poetry install --no-root
-cp .env_template .env
-```
-
-Then fill `.env` with the omitted values (ask in Zulip). At the moment of this writing, the only value omitted is `AUTH_FIEF_CLIENT_SECRET`.
-
-### Start the API/Backend
-
-Start the background worker:
-
-```bash
-poetry run celery -A reflector.worker.app worker --loglevel=info
-```
-
-### Redis (Mac)
-
-```bash
-yarn add redis
-poetry run celery -A reflector.worker.app worker --loglevel=info
-redis-server
-```
-
-### Redis (Windows)
-
-**Option 1**
+**Run in development mode**
 
 ```bash
 docker compose up -d redis
+
+# on the first run, or if the schemas changed
+uv run alembic upgrade head
+
+# start the worker
+uv run celery -A reflector.worker.app worker --loglevel=info
+
+# start the app
+uv run -m reflector.app --reload
 ```
 
-**Option 2**
+Then fill `.env` with the omitted values (ask in Zulip).
 
-Install:
-
-- [Git for Windows](https://gitforwindows.org/)
-- [Windows Subsystem for Linux (WSL)](https://docs.microsoft.com/en-us/windows/wsl/install)
-- Install your preferred Linux distribution via the Microsoft Store (e.g., Ubuntu).
-
-Open your Linux distribution and update the package list:
-
-```bash
-sudo apt update
-sudo apt install redis-server
-redis-server
-```
-
-## Update the database schema (run on first install, and after each pull containing a migration)
-
-```bash
-poetry run alembic heads
-```
-
-## Main Server
-
-```bash
-poetry run python -m reflector.app
-```
-
-### Crontab (optional)
+**Crontab (optional)**
 
 For crontab (only healthcheck for now), start the celery beat (you don't need it on your local dev environment):
 
 ```bash
-poetry run celery -A reflector.worker.app beat
+uv run celery -A reflector.worker.app beat
 ```
 
-#### Using docker
+### GPU models
 
-Use:
+Currently, reflector heavily use custom local models, deployed on modal. All the micro services are available in server/gpu/
 
-```bash
-docker-compose up server
-```
-
-### Using local GPT4All
-
-- Start GPT4All with any model you want
-- Ensure the API server is activated in GPT4all
-- Run with: `LLM_BACKEND=openai LLM_URL=http://localhost:4891/v1/completions LLM_OPENAI_MODEL="GPT4All Falcon" python -m reflector.app`
-
-### Using local files
-
-```
-poetry run python -m reflector.tools.process path/to/audio.wav
-```
-
-## AI Models
-
-### Modal
-To deploy llm changes to modal, you need.
+To deploy llm changes to modal, you need:
 - a modal account
 - set up the required secret in your modal account (REFLECTOR_GPU_APIKEY)
 - install the modal cli
 - connect your modal cli to your account if not done previously
 - `modal run path/to/required/llm`
 
-_(Documentation for this section is pending.)_
+## Using local files
+
+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:
@@ -36,13 +39,31 @@ services:
     ports:
       - 6379:6379
   web:
-    image: node:18
+    image: node:22-alpine
     ports:
       - "3000:3000"
-    command: sh -c "yarn install && yarn dev"
+    command: sh -c "corepack enable && pnpm install && pnpm dev"
     restart: unless-stopped
     working_dir: /app
     volumes:
       - ./www:/app/
+      - /app/node_modules
     env_file:
       - ./www/.env.local
+    environment:
+      - NODE_ENV=development
+
+  postgres:
+    image: postgres:17
+    ports:
+      - 5432:5432
+    environment:
+      POSTGRES_USER: reflector
+      POSTGRES_PASSWORD: reflector
+      POSTGRES_DB: reflector
+    volumes:
+      - ./data/postgres:/var/lib/postgresql/data
+
+networks:
+  default:
+    attachable: true

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/README.md

@@ -0,0 +1,171 @@
+# Reflector GPU implementation - Transcription and LLM
+
+This repository hold an API for the GPU implementation of the Reflector API service,
+and use [Modal.com](https://modal.com)
+
+- `reflector_diarizer.py` - Diarization API
+- `reflector_transcriber.py` - Transcription API (Whisper)
+- `reflector_transcriber_parakeet.py` - Transcription API (NVIDIA Parakeet)
+- `reflector_translator.py` - Translation API
+
+## Modal.com deployment
+
+Create a modal secret, and name it `reflector-gpu`.
+It should contain an `REFLECTOR_APIKEY` environment variable with a value.
+
+The deployment is done using [Modal.com](https://modal.com) service.
+
+```
+$ modal deploy reflector_transcriber.py
+...
+└── 🔨 Created web => https://xxxx--reflector-transcriber-web.modal.run
+
+$ modal deploy reflector_transcriber_parakeet.py
+...
+└── 🔨 Created web => https://xxxx--reflector-transcriber-parakeet-web.modal.run
+
+$ modal deploy reflector_llm.py
+...
+└── 🔨 Created web => https://xxxx--reflector-llm-web.modal.run
+```
+
+Then in your reflector api configuration `.env`, you can set these keys:
+
+```
+TRANSCRIPT_BACKEND=modal
+TRANSCRIPT_URL=https://xxxx--reflector-transcriber-web.modal.run
+TRANSCRIPT_MODAL_API_KEY=REFLECTOR_APIKEY
+
+DIARIZATION_BACKEND=modal
+DIARIZATION_URL=https://xxxx--reflector-diarizer-web.modal.run
+DIARIZATION_MODAL_API_KEY=REFLECTOR_APIKEY
+
+TRANSLATION_BACKEND=modal
+TRANSLATION_URL=https://xxxx--reflector-translator-web.modal.run
+TRANSLATION_MODAL_API_KEY=REFLECTOR_APIKEY
+```
+
+## API
+
+Authentication must be passed with the `Authorization` header, using the `bearer` scheme.
+
+```
+Authorization: bearer <REFLECTOR_APIKEY>
+```
+
+### LLM
+
+`POST /llm`
+
+**request**
+```
+{
+    "prompt": "xxx"
+}
+```
+
+**response**
+```
+{
+    "text": "xxx completed"
+}
+```
+
+### Transcription
+
+#### Parakeet Transcriber (`reflector_transcriber_parakeet.py`)
+
+NVIDIA Parakeet is a state-of-the-art ASR model optimized for real-time transcription with superior word-level timestamps.
+
+**GPU Configuration:**
+- **A10G GPU** - Used for `/v1/audio/transcriptions` endpoint (small files, live transcription)
+  - Higher concurrency (max_inputs=10)
+  - Optimized for multiple small audio files
+  - Supports batch processing for efficiency
+
+- **L40S GPU** - Used for `/v1/audio/transcriptions-from-url` endpoint (large files)
+  - Lower concurrency but more powerful processing
+  - Optimized for single large audio files
+  - VAD-based chunking for long-form audio
+
+##### `/v1/audio/transcriptions` - Small file transcription
+
+**request** (multipart/form-data)
+- `file` or `files[]` - audio file(s) to transcribe
+- `model` - model name (default: `nvidia/parakeet-tdt-0.6b-v2`)
+- `language` - language code (default: `en`)
+- `batch` - whether to use batch processing for multiple files (default: `true`)
+
+**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"
+}
+```
+
+For multiple files with batch=true:
+```json
+{
+    "results": [
+        {
+            "filename": "audio1.mp3",
+            "text": "transcribed text",
+            "words": [...]
+        },
+        {
+            "filename": "audio2.mp3",
+            "text": "transcribed text",
+            "words": [...]
+        }
+    ]
+}
+```
+
+##### `/v1/audio/transcriptions-from-url` - Large file transcription
+
+**request** (application/json)
+```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 from large file",
+    "words": [
+        {"word": "hello", "start": 0.0, "end": 0.5},
+        {"word": "world", "start": 0.5, "end": 1.0}
+    ]
+}
+```
+
+**Supported file types:** mp3, mp4, mpeg, mpga, m4a, wav, webm
+
+#### Whisper Transcriber (`reflector_transcriber.py`)
+
+`POST /transcribe`
+
+**request** (multipart/form-data)
+
+- `file` - audio file
+- `language` - language code (e.g. `en`)
+
+**response**
+```
+{
+    "text": "xxx",
+    "words": [
+        {"text": "xxx", "start": 0.0, "end": 1.0}
+    ]
+}
+```

gpu/modal_deployments/reflector_diarizer.py

@@ -0,0 +1,253 @@
+"""
+Reflector GPU backend - diarizer
+===================================
+"""
+
+import os
+import uuid
+from typing import Mapping, NewType
+from urllib.parse import urlparse
+
+import modal
+
+PYANNOTE_MODEL_NAME: str = "pyannote/speaker-diarization-3.1"
+MODEL_DIR = "/root/diarization_models"
+UPLOADS_PATH = "/uploads"
+SUPPORTED_FILE_EXTENSIONS = ["mp3", "mp4", "mpeg", "mpga", "m4a", "wav", "webm"]
+
+DiarizerUniqFilename = NewType("DiarizerUniqFilename", str)
+AudioFileExtension = NewType("AudioFileExtension", str)
+
+app = modal.App(name="reflector-diarizer")
+
+# Volume for temporary file uploads
+upload_volume = modal.Volume.from_name("diarizer-uploads", create_if_missing=True)
+
+
+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[DiarizerUniqFilename, AudioFileExtension]:
+    import requests
+    from fastapi import HTTPException
+
+    print(f"Checking audio file at: {audio_file_url}")
+    response = requests.head(audio_file_url, allow_redirects=True)
+    if response.status_code == 404:
+        raise HTTPException(status_code=404, detail="Audio file not found")
+
+    print(f"Downloading audio file from: {audio_file_url}")
+    response = requests.get(audio_file_url, allow_redirects=True)
+
+    if response.status_code != 200:
+        print(f"Download failed with status {response.status_code}: {response.text}")
+        raise HTTPException(
+            status_code=response.status_code,
+            detail=f"Failed to download audio file: {response.status_code}",
+        )
+
+    audio_suffix = detect_audio_format(audio_file_url, response.headers)
+    unique_filename = DiarizerUniqFilename(f"{uuid.uuid4()}.{audio_suffix}")
+    file_path = f"{UPLOADS_PATH}/{unique_filename}"
+
+    print(f"Writing file to: {file_path} (size: {len(response.content)} bytes)")
+    with open(file_path, "wb") as f:
+        f.write(response.content)
+
+    upload_volume.commit()
+    print(f"File saved as: {unique_filename}")
+    return unique_filename, audio_suffix
+
+
+def migrate_cache_llm():
+    """
+    XXX The cache for model files in Transformers v4.22.0 has been updated.
+    Migrating your old cache. This is a one-time only operation. You can
+    interrupt this and resume the migration later on by calling
+    `transformers.utils.move_cache()`.
+    """
+    from transformers.utils.hub import move_cache
+
+    print("Moving LLM cache")
+    move_cache(cache_dir=MODEL_DIR, new_cache_dir=MODEL_DIR)
+    print("LLM cache moved")
+
+
+def download_pyannote_audio():
+    from pyannote.audio import Pipeline
+
+    Pipeline.from_pretrained(
+        PYANNOTE_MODEL_NAME,
+        cache_dir=MODEL_DIR,
+        use_auth_token=os.environ["HF_TOKEN"],
+    )
+
+
+diarizer_image = (
+    modal.Image.debian_slim(python_version="3.10.8")
+    .pip_install(
+        "pyannote.audio==3.1.0",
+        "requests",
+        "onnx",
+        "torchaudio",
+        "onnxruntime-gpu",
+        "torch==2.0.0",
+        "transformers==4.34.0",
+        "sentencepiece",
+        "protobuf",
+        "numpy",
+        "huggingface_hub",
+        "hf-transfer",
+    )
+    .run_function(
+        download_pyannote_audio,
+        secrets=[modal.Secret.from_name("hf_token")],
+    )
+    .run_function(migrate_cache_llm)
+    .env(
+        {
+            "LD_LIBRARY_PATH": (
+                "/usr/local/lib/python3.10/site-packages/nvidia/cudnn/lib/:"
+                "/opt/conda/lib/python3.10/site-packages/nvidia/cublas/lib/"
+            )
+        }
+    )
+)
+
+
+@app.cls(
+    gpu="A100",
+    timeout=60 * 30,
+    image=diarizer_image,
+    volumes={UPLOADS_PATH: upload_volume},
+    enable_memory_snapshot=True,
+    experimental_options={"enable_gpu_snapshot": True},
+    secrets=[
+        modal.Secret.from_name("hf_token"),
+    ],
+)
+@modal.concurrent(max_inputs=1)
+class Diarizer:
+    @modal.enter(snap=True)
+    def enter(self):
+        import torch
+        from pyannote.audio import Pipeline
+
+        self.use_gpu = torch.cuda.is_available()
+        self.device = "cuda" if self.use_gpu else "cpu"
+        print(f"Using device: {self.device}")
+        self.diarization_pipeline = Pipeline.from_pretrained(
+            PYANNOTE_MODEL_NAME,
+            cache_dir=MODEL_DIR,
+            use_auth_token=os.environ["HF_TOKEN"],
+        )
+        self.diarization_pipeline.to(torch.device(self.device))
+
+    @modal.method()
+    def diarize(self, filename: str, timestamp: float = 0.0):
+        import torchaudio
+
+        upload_volume.reload()
+
+        file_path = f"{UPLOADS_PATH}/{filename}"
+        if not os.path.exists(file_path):
+            raise FileNotFoundError(f"File not found: {file_path}")
+
+        print(f"Diarizing audio from: {file_path}")
+        waveform, sample_rate = torchaudio.load(file_path)
+        diarization = self.diarization_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:]),
+                }
+            )
+        print("Diarization complete")
+        return {"diarization": words}
+
+
+# -------------------------------------------------------------------
+# Web API
+# -------------------------------------------------------------------
+
+
+@app.function(
+    timeout=60 * 10,
+    scaledown_window=60 * 3,
+    secrets=[
+        modal.Secret.from_name("reflector-gpu"),
+    ],
+    volumes={UPLOADS_PATH: upload_volume},
+    image=diarizer_image,
+)
+@modal.concurrent(max_inputs=40)
+@modal.asgi_app()
+def web():
+    from fastapi import Depends, FastAPI, HTTPException, status
+    from fastapi.security import OAuth2PasswordBearer
+    from pydantic import BaseModel
+
+    diarizerstub = Diarizer()
+
+    app = FastAPI()
+
+    oauth2_scheme = OAuth2PasswordBearer(tokenUrl="token")
+
+    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 DiarizationResponse(BaseModel):
+        result: dict
+
+    @app.post("/diarize", dependencies=[Depends(apikey_auth)])
+    def diarize(audio_file_url: str, timestamp: float = 0.0) -> DiarizationResponse:
+        unique_filename, audio_suffix = download_audio_to_volume(audio_file_url)
+
+        try:
+            func = diarizerstub.diarize.spawn(
+                filename=unique_filename, timestamp=timestamp
+            )
+            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

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

@@ -0,0 +1,658 @@
+import logging
+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 = "nvidia/parakeet-tdt-0.6b-v2"
+SUPPORTED_FILE_EXTENSIONS = ["mp3", "mp4", "mpeg", "mpga", "m4a", "wav", "webm"]
+SAMPLERATE = 16000
+UPLOADS_PATH = "/uploads"
+CACHE_PATH = "/cache"
+VAD_CONFIG = {
+    "batch_max_duration": 30.0,
+    "silence_padding": 0.5,
+    "window_size": 512,
+}
+
+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
+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]==2.5.0",
+        "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: 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"]
+        ]
+
+        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: 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"]
+            ]
+
+            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,
+        ) -> 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"]
+            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)
+
+                    yield TimeSegment(start_time, end_time)
+                    start = None
+
+            vad_iterator.reset_states()
+
+        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]
+
+                                  ↓ (max_duration=10)
+
+              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
+
+                total_duration = end_time - batch_start_time
+
+                if total_duration <= max_duration:
+                    batch_end_time = end_time
+                    continue
+
+                yield TimeSegment(batch_start_time, batch_end_time)
+                batch_start_time = start_time
+                batch_end_time = end_time
+
+            if batch_start_time is None or batch_end_time is None:
+                return
+
+            yield TimeSegment(batch_start_time, batch_end_time)
+
+        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.
+
+            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]
+
+                padded_segment = pad_audio(audio_segment, SAMPLERATE)
+
+                yield AudioSegment(start_time, end_time, padded_segment)
+
+        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: 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, segment) in enumerate(zip(results, segments_info)):
+                start_time, end_time = segment.start, segment.end
+                text = output.text.strip()
+                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),
+                    )
+                    for word_info in output.timestamp["word"]
+                ]
+
+                yield TranscriptResult(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)
+
+        all_text_parts: list[str] = []
+        all_words: list[WordTiming] = []
+
+        raw_segments = vad_segment_generator(audio_array)
+        speech_segments = batch_speech_segments(
+            raw_segments,
+            VAD_CONFIG["batch_max_duration"],
+        )
+        audio_segments = batch_segment_to_audio_segment(speech_segments, audio_array)
+
+        for batch in audio_segments:
+            audio_segment = batch.audio
+            results = transcribe_batch(self.model, [audio_segment])
+
+            for result in emit_results(
+                results,
+                [batch],
+            ):
+                if not result.text:
+                    continue
+                all_text_parts.append(result.text)
+                all_words.extend(result.words)
+
+        all_words = enforce_word_timing_constraints(all_words)
+
+        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()

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/.env_template

@@ -1,21 +0,0 @@
-TRANSCRIPT_BACKEND=modal
-TRANSCRIPT_URL=https://monadical-sas--reflector-transcriber-web.modal.run
-TRANSCRIPT_MODAL_API_KEY=***REMOVED***
-
-LLM_BACKEND=modal
-LLM_URL=https://monadical-sas--reflector-llm-web.modal.run
-LLM_MODAL_API_KEY=***REMOVED***
-
-AUTH_BACKEND=fief
-AUTH_FIEF_URL=https://auth.reflector.media/reflector-local
-AUTH_FIEF_CLIENT_ID=***REMOVED***
-AUTH_FIEF_CLIENT_SECRET=<ask in zulip> <-----------------------------------------------------------------------------------------
-
-TRANSLATE_URL=https://monadical-sas--reflector-translator-web.modal.run
-ZEPHYR_LLM_URL=https://monadical-sas--reflector-llm-zephyr-web.modal.run
-DIARIZATION_URL=https://monadical-sas--reflector-diarizer-web.modal.run
-
-BASE_URL=https://xxxxx.ngrok.app
-DIARIZATION_ENABLED=false
-
-SQS_POLLING_TIMEOUT_SECONDS=60

server/.gitignore

@@ -176,7 +176,9 @@ artefacts/
 audio_*.wav
 
 # ignore local database
-reflector.sqlite3
+*.sqlite3
+*.db
 data/
 
 dump.rdb
+

server/.python-version

@@ -1 +1 @@
-3.11.6
+3.12

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

@@ -1,30 +1,41 @@
-FROM python:3.11-slim as base
+FROM python:3.12-slim
 
-ENV PIP_DEFAULT_TIMEOUT=100 \
-    PIP_DISABLE_PIP_VERSION_CHECK=1 \
-    PIP_NO_CACHE_DIR=1 \
-    PYTHONDONTWRITEBYTECODE=1 \
-    PYTHONUNBUFFERED=1 \
-    POETRY_VERSION=1.3.1
+ENV PYTHONUNBUFFERED=1 \
+    UV_LINK_MODE=copy \
+    UV_NO_CACHE=1
 
 # builder install base dependencies
-FROM base AS builder
 WORKDIR /tmp
-RUN pip install "poetry==$POETRY_VERSION"
-RUN python -m venv /venv
+RUN apt-get update && apt-get install -y curl ffmpeg && apt-get clean
+ADD https://astral.sh/uv/install.sh /uv-installer.sh
+RUN sh /uv-installer.sh && rm /uv-installer.sh
+ENV PATH="/root/.local/bin/:$PATH"
 
 # install application dependencies
-COPY pyproject.toml poetry.lock /tmp
-RUN . /venv/bin/activate && poetry config virtualenvs.create false
-RUN . /venv/bin/activate && poetry install --only main,aws --no-root --no-interaction --no-ansi
+RUN mkdir -p /app
+WORKDIR /app
+COPY pyproject.toml uv.lock README.md /app/
+RUN uv sync --compile-bytecode --locked
+
+# pre-download nltk packages
+RUN uv run python -c "import nltk; nltk.download('punkt_tab'); nltk.download('averaged_perceptron_tagger_eng')"
 
 # bootstrap
-FROM base AS final
-COPY --from=builder /venv /venv
-RUN mkdir -p /app
-COPY reflector /app/reflector
-COPY migrations /app/migrations
-COPY images /app/images
 COPY alembic.ini runserver.sh /app/
+COPY images /app/images
+COPY migrations /app/migrations
+COPY reflector /app/reflector
 WORKDIR /app
+
+# Create symlink for libgomp if it doesn't exist (for ARM64 compatibility)
+RUN if [ "$(uname -m)" = "aarch64" ] && [ ! -f /usr/lib/libgomp.so.1 ]; then \
+    LIBGOMP_PATH=$(find /app/.venv/lib -path "*/torch.libs/libgomp*.so.*" 2>/dev/null | head -n1); \
+    if [ -n "$LIBGOMP_PATH" ]; then \
+    ln -sf "$LIBGOMP_PATH" /usr/lib/libgomp.so.1; \
+    fi \
+    fi
+
+# Pre-check just to make sure the image will not fail
+RUN uv run python -c "import silero_vad.model"
+
 CMD ["./runserver.sh"]

server/README.md

@@ -20,3 +20,25 @@ Polls SQS every 60 seconds via /server/reflector/worker/process.py:24-62:
 # Every 60 seconds, check for new recordings
 sqs = boto3.client("sqs", ...)
 response = sqs.receive_message(QueueUrl=queue_url, ...)
+
+# Requeue
+
+```bash
+uv run /app/requeue_uploaded_file.py TRANSCRIPT_ID
+```
+
+## Pipeline Management
+
+### Continue stuck pipeline from final summaries (identify_participants) step:
+
+```bash
+uv run python -c "from reflector.pipelines.main_live_pipeline import task_pipeline_final_summaries; result = task_pipeline_final_summaries.delay(transcript_id='TRANSCRIPT_ID'); print(f'Task queued: {result.id}')"
+```
+
+### Run full post-processing pipeline (continues to completion):
+
+```bash
+uv run python -c "from reflector.pipelines.main_live_pipeline import pipeline_post; pipeline_post(transcript_id='TRANSCRIPT_ID')"
+```
+
+.

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

@@ -7,11 +7,9 @@
 ## User authentication
 ## =======================================================
 
-## Using fief (fief.dev)
-AUTH_BACKEND=fief
-AUTH_FIEF_URL=https://auth.reflector.media/reflector-local
-AUTH_FIEF_CLIENT_ID=***REMOVED***
-AUTH_FIEF_CLIENT_SECRET=<ask in zulip>
+## Using jwt/authentik
+AUTH_BACKEND=jwt
+AUTH_JWT_AUDIENCE=
 
 ## =======================================================
 ## Transcription backend
@@ -22,24 +20,24 @@ AUTH_FIEF_CLIENT_SECRET=<ask in zulip>
 
 ## Using local whisper
 #TRANSCRIPT_BACKEND=whisper
-#WHISPER_MODEL_SIZE=tiny
 
 ## Using serverless modal.com (require reflector-gpu-modal deployed)
 #TRANSCRIPT_BACKEND=modal
 #TRANSCRIPT_URL=https://xxxxx--reflector-transcriber-web.modal.run
-#TRANSLATE_URL=https://xxxxx--reflector-translator-web.modal.run
 #TRANSCRIPT_MODAL_API_KEY=xxxxx
 
 TRANSCRIPT_BACKEND=modal
-TRANSCRIPT_URL=https://monadical-sas--reflector-transcriber-web.modal.run
-TRANSCRIPT_MODAL_API_KEY=***REMOVED***
+TRANSCRIPT_URL=https://monadical-sas--reflector-transcriber-parakeet-web.modal.run
+TRANSCRIPT_MODAL_API_KEY=
 
 ## =======================================================
-## Transcription backend
+## Translation backend
 ##
 ## Only available in modal atm
 ## =======================================================
+TRANSLATION_BACKEND=modal
 TRANSLATE_URL=https://monadical-sas--reflector-translator-web.modal.run
+#TRANSLATION_MODAL_API_KEY=xxxxx
 
 ## =======================================================
 ## LLM backend
@@ -49,28 +47,11 @@ TRANSLATE_URL=https://monadical-sas--reflector-translator-web.modal.run
 ## llm backend implementation
 ## =======================================================
 
-## Using serverless modal.com (require reflector-gpu-modal deployed)
-LLM_BACKEND=modal
-LLM_URL=https://monadical-sas--reflector-llm-web.modal.run
-LLM_MODAL_API_KEY=***REMOVED***
-ZEPHYR_LLM_URL=https://monadical-sas--reflector-llm-zephyr-web.modal.run
-
-
-## Using OpenAI
-#LLM_BACKEND=openai
-#LLM_OPENAI_KEY=xxx
-#LLM_OPENAI_MODEL=gpt-3.5-turbo
-
-## Using GPT4ALL
-#LLM_BACKEND=openai
-#LLM_URL=http://localhost:4891/v1/completions
-#LLM_OPENAI_MODEL="GPT4All Falcon"
-
-## Default LLM MODEL NAME
-#DEFAULT_LLM=lmsys/vicuna-13b-v1.5
-
-## Cache directory to store models
-CACHE_DIR=data
+## Context size for summary generation (tokens)
+# LLM_MODEL=microsoft/phi-4
+LLM_CONTEXT_WINDOW=16000
+LLM_URL=
+LLM_API_KEY=sk-
 
 ## =======================================================
 ## Diarization
@@ -79,7 +60,9 @@ CACHE_DIR=data
 ## To allow diarization, you need to expose expose the files to be dowloded by the pipeline
 ## =======================================================
 DIARIZATION_ENABLED=false
+DIARIZATION_BACKEND=modal
 DIARIZATION_URL=https://monadical-sas--reflector-diarizer-web.modal.run
+#DIARIZATION_MODAL_API_KEY=xxxxx
 
 
 ## =======================================================
@@ -89,3 +72,26 @@ 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/evaluate/evaluate_transcription.py

@@ -1,204 +0,0 @@
-import re
-from pathlib import Path
-from typing import Any, List
-
-from jiwer import wer
-from Levenshtein import distance
-from pydantic import BaseModel, Field, field_validator
-from tqdm.auto import tqdm
-from whisper.normalizers import EnglishTextNormalizer
-
-
-class EvaluationResult(BaseModel):
-    """
-    Result object of the model evaluation
-    """
-    accuracy: float = Field(default=0.0)
-    total_test_samples: int = Field(default=0)
-
-
-class EvaluationTestSample(BaseModel):
-    """
-    Represents one test sample
-    """
-
-    reference_text: str
-    predicted_text: str
-
-    def update(self, reference_text:str, predicted_text:str) -> None:
-        self.reference_text = reference_text
-        self.predicted_text = predicted_text
-
-
-class TestDatasetLoader(BaseModel):
-    """
-    Test samples loader
-    """
-
-    test_dir: Path = Field(default=Path(__file__).parent)
-    total_samples: int = Field(default=0)
-
-    @field_validator("test_dir")
-    def validate_file_path(cls, path):
-        """
-        Check the file path
-        """
-        if not path.exists():
-            raise ValueError("Path does not exist")
-        return path
-
-    def _load_test_data(self) -> tuple[Path, Path]:
-        """
-        Loader function to validate input files and generate samples
-        """
-        PREDICTED_TEST_SAMPLES_DIR = self.test_dir / "predicted_texts"
-        REFERENCE_TEST_SAMPLES_DIR = self.test_dir / "reference_texts"
-
-        for filename in PREDICTED_TEST_SAMPLES_DIR.iterdir():
-            match = re.search(r"(\d+)\.txt$", filename.as_posix())
-            if match:
-                sample_id = match.group(1)
-                pred_file_path = PREDICTED_TEST_SAMPLES_DIR / filename
-                ref_file_name = "ref_sample_" + str(sample_id) + ".txt"
-                ref_file_path = REFERENCE_TEST_SAMPLES_DIR / ref_file_name
-                if ref_file_path.exists():
-                    self.total_samples += 1
-                    yield ref_file_path, pred_file_path
-
-    def __iter__(self) -> EvaluationTestSample:
-        """
-        Iter method for the test loader
-        """
-        for pred_file_path, ref_file_path in self._load_test_data():
-            with open(pred_file_path, "r", encoding="utf-8") as file:
-                pred_text = file.read()
-            with open(ref_file_path, "r", encoding="utf-8") as file:
-                ref_text = file.read()
-            yield EvaluationTestSample(reference_text=ref_text, predicted_text=pred_text)
-
-
-class EvaluationConfig(BaseModel):
-    """
-    Model for evaluation parameters
-    """
-    insertion_penalty: int = Field(default=1)
-    substitution_penalty: int = Field(default=1)
-    deletion_penalty: int = Field(default=1)
-    normalizer: Any = Field(default=EnglishTextNormalizer())
-    test_directory: str = Field(default=str(Path(__file__).parent))
-
-
-class ModelEvaluator:
-    """
-    Class that comprises all model evaluation related processes and methods
-    """
-
-    # The 2 popular methods of WER differ slightly. More dimensions of accuracy
-    # will be added. For now, the average of these 2 will serve as the metric.
-    WEIGHTED_WER_LEVENSHTEIN = 0.0
-    WER_LEVENSHTEIN = []
-    WEIGHTED_WER_JIWER = 0.0
-    WER_JIWER = []
-
-    evaluation_result = EvaluationResult()
-    test_dataset_loader = None
-    evaluation_config = None
-
-    def __init__(self, **kwargs):
-        self.evaluation_config = EvaluationConfig(**kwargs)
-        self.test_dataset_loader = TestDatasetLoader(test_dir=self.evaluation_config.test_directory)
-
-    def __repr__(self):
-        return f"ModelEvaluator({self.evaluation_config})"
-
-    def describe(self) -> dict:
-        """
-        Returns the parameters defining the evaluator
-        """
-        return self.evaluation_config.model_dump()
-
-    def _normalize(self, sample: EvaluationTestSample) -> None:
-        """
-        Normalize both reference and predicted text
-        """
-        sample.update(
-            self.evaluation_config.normalizer(sample.reference_text),
-            self.evaluation_config.normalizer(sample.predicted_text),
-        )
-
-    def _calculate_wer(self, sample: EvaluationTestSample) -> float:
-        """
-        Based on weights for (insert, delete, substitute), calculate
-        the Word Error Rate
-        """
-        levenshtein_distance = distance(
-            s1=sample.reference_text,
-            s2=sample.predicted_text,
-            weights=(
-                self.evaluation_config.insertion_penalty,
-                self.evaluation_config.deletion_penalty,
-                self.evaluation_config.substitution_penalty,
-            ),
-        )
-        wer = levenshtein_distance / len(sample.reference_text)
-        return wer
-
-    def _calculate_wers(self) -> None:
-        """
-        Compute WER
-        """
-        for sample in tqdm(self.test_dataset_loader, desc="Evaluating"):
-            self._normalize(sample)
-            wer_item_l = {
-                "wer": self._calculate_wer(sample),
-                "no_of_words": len(sample.reference_text),
-            }
-            wer_item_j = {
-                "wer": wer(sample.reference_text, sample.predicted_text),
-                "no_of_words": len(sample.reference_text),
-            }
-            self.WER_LEVENSHTEIN.append(wer_item_l)
-            self.WER_JIWER.append(wer_item_j)
-
-    def _calculate_weighted_wer(self, wers: List[float]) -> float:
-        """
-        Calculate the weighted WER from WER
-        """
-        total_wer = 0.0
-        total_words = 0.0
-        for item in wers:
-            total_wer += item["no_of_words"] * item["wer"]
-            total_words += item["no_of_words"]
-        return total_wer / total_words
-
-    def _calculate_model_accuracy(self) -> None:
-        """
-        Compute model accuracy
-        """
-        self._calculate_wers()
-        weighted_wer_levenshtein = self._calculate_weighted_wer(self.WER_LEVENSHTEIN)
-        weighted_wer_jiwer = self._calculate_weighted_wer(self.WER_JIWER)
-
-        final_weighted_wer = (weighted_wer_levenshtein + weighted_wer_jiwer) / 2
-        self.evaluation_result.accuracy = (1 - final_weighted_wer) * 100
-
-    def evaluate(self, recalculate: bool = False) -> EvaluationResult:
-        """
-        Triggers the model evaluation
-        """
-        if not self.evaluation_result.accuracy or recalculate:
-            self._calculate_model_accuracy()
-        return EvaluationResult(
-                accuracy=self.evaluation_result.accuracy,
-                total_test_samples=self.test_dataset_loader.total_samples
-        )
-
-
-eval_config = {"insertion_penalty": 1, "deletion_penalty": 2, "substitution_penalty": 1}
-
-evaluator = ModelEvaluator(**eval_config)
-evaluation = evaluator.evaluate()
-
-print(evaluator)
-print(evaluation)
-print("Model accuracy : {:.2f} %".format(evaluation.accuracy))

server/evaluate/reference_texts/ref_sample_2.txt

@@ -1,620 +0,0 @@
-Technologies ticker symbol w-e-l-l on
-
-the TSX recently reported its 2023 q1
-
-results beating the streets consensus
-
-estimate for revenue and adjusted ebitda
-
-and in a report issued this week Raymond
-
-James analyst said quote we're impressed
-
-by Wells capacity to drive powerful
-
-growth across its diverse business units
-
-in the absence of M A joining me today
-
-is CEO Hamed chabazi to look at what's
-
-next for well health good to see you sir
-
-how are you great to see you Richard
-
-thanks very much for having me great to
-
-have you uh congratulations on your 17th
-
-consecutive quarter of record Revenue
-
-can you share some insights into what's
-
-Driven these results historically and in
-
-the past quarter as well
-
-yeah thank you we we're very excited
-
-about our uh q1 2023 results and as you
-
-mentioned uh we've had a long you know
-
-successful uh string of of uh you know
-
-continued growth and record growth
-
-um we also had accelerating organic
-
-growth and I think um a big part of the
-
-success of our franchise here is the
-
-incredibly sticky and predictable
-
-Revenue that we have you know well over
-
-90 of our business is either highly
-
-reoccurring as in uh the you know highly
-
-predictable uh results of our two-sided
-
-network of patients and providers or
-
-truly recurring as in scheduled or
-
-subscribed revenues and this allows us
-
-to essentially make sure that that uh
-
-you know we're on track it obviously you
-
-know like any other business things
-
-happen uh and sometimes it's hard to
-
-meet those results but what's really
-
-being unique about our platform is we do
-
-have exposure to all kinds of different
-
-aspects of healthcare you know we have
-
-Prime primary care and Specialized Care
-
-on both sides of the Border in the US
-
-and Canada so we have exposure to
-
-different types of business models we
-
-have exposure to the U.S payer Network
-
-which has higher per unit economics than
-
-Canada and of course the stability and
-
-uh and and sort of higher Fidelity uh
-
-kind of Collections and revenue cycle
-
-process that Canada has over the United
-
-States where you don't have to kind of
-
-deal with all of that uh at that payment
-
-noise so just a lot of I think strength
-
-built into the platform because of the
-
-diversity of different Healthcare
-
-businesses that we support
-
-and uh where do you see Well's future
-
-growth coming from which part of the
-
-business uh excites you the most right
-
-now yeah well look the centrifugal force
-
-of well is the healthcare provider and
-
-we exist to uh Tech enable and
-
-ameliorate the business of that of that
-
-Tech of that healthcare provider uh and
-
-and and that's what we're laser focused
-
-on and and what we're seeing is
-
-providers not wanting to run businesses
-
-anymore it's very simple and so we have
-
-a digital platform and providers can
-
-either acquire what they want and need
-
-from our digital platform and implement
-
-it themselves
-
-or they can decide that they don't want
-
-to run a business anymore they don't
-
-want to configure and manage technology
-
-which is becoming a bigger and bigger
-
-part of their world every single day and
-
-when we see what we've seen with that
-
-Dynamic is that uh is that a lot of them
-
-are now just wanting to work in a place
-
-where where all the technology is
-
-configured for them it's wrapped around
-
-them and they have a competent operating
-
-partner that is supporting the organ the
-
-the practice uh and and taking care of
-
-the front office in the back office so
-
-that they can focus on providing care
-
-this results in them seeing more
-
-patients uh and and being happier
-
-because you know they became doctors to
-
-see patients not so they can manage uh
-
-workers and and deal with HR issues and
-
-deal with labs and all that kind of
-
-stuff excellent and I know too that
-
-Acquisitions have played a key role in
-
-well can you share any insights into how
-
-the Acquisitions fit into Wells growth
-
-strategy
-
-sure in in look in 2020 and 2021 we did
-
-a lot of Acquisitions in 2022 we took a
-
-bit of a breather and we've really
-
-focused on integration and I think
-
-that's one of the reasons why you saw
-
-this accelerating organic growth we
-
-really were able to demonstrate that we
-
-could bring together the different
-
-elements of our technology platform we
-
-started to sell bundles we started to
-
-really derive Synergy uh and activate uh
-
-you know more sales as a result of
-
-selling uh all the different products
-
-and services with one voice with One
-
-Vision uh so we made it easier for
-
-providers to use their technology and I
-
-think that was a big reason uh for our
-
-growth now M A as you know where Capital
-
-allocation company we're never far from
-
-it and so we did continue to have you
-
-know tuck-ins here and there and in fact
-
-today uh we announced that we've
-
-acquired uh the Alberta operations of uh
-
-MCI one Health and other publicly traded
-
-company uh who was looking to raise
-
-funds to support their business we're
-
-very pleased with with this acquisition
-
-it just demonstrates our continued
-
-discipline these are you know great
-
-primary care clinics in in Canada right
-
-in the greater Calgary area and uh uh
-
-you know just allows us to grow our
-
-footprint in Alberta which is an
-
-important Province for us and it it's
-
-it's if you look at the price if you
-
-look at what we're getting uh you know
-
-it's just demonstrative of our continued
-
-uh discipline and just you know a few
-
-days ago at our conference call I
-
-mentioned uh that we had you know a
-
-really strong lineup of Acquisitions uh
-
-and you know they're starting to uh uh I
-
-think uh come to fruition for us
-
-a company on the grown-up question I you
-
-recently announced a new AI investment
-
-program last month what specific areas
-
-of healthcare technology or AI are you
-
-focusing on and what's the strategy when
-
-it comes to AI
-
-yes uh look AI as as I'm sure you're
-
-aware is it's become you know really uh
-
-an incredibly important topic in in all
-
-aspects of of business and and you know
-
-not just business socially as well
-
-everyone's talking about uh this this
-
-new breakthrough disruptive technology
-
-the large language models and generative
-
-AI
-
-um I mean look AI uh has been about a 80
-
-year old overnight success a lot of
-
-people have been working on this for a
-
-long time generative AI is just sort of
-
-you know the culmination of a lot of
-
-things coming together and working uh
-
-but it is uncorked enormous uh
-
-Innovation and and we think that um this
-
-there's a very good news story about
-
-this in healthcare particularly where we
-
-were looking to look we were looking to
-
-unlock uh the value of of the data that
-
-that we all produce every single day
-
-um as as humans and and so we've
-
-established an AI investment program
-
-because no one company can can tackle
-
-all of these Innovations themselves and
-
-what well has done too is it's taken a
-
-very much an ecosystem approach by
-
-establishing its apps.health Marketplace
-
-and so we're very excited about not only
-
-uh allocating Capital into promising
-
-young AI companies that are focused on
-
-digital health and solving Healthcare
-
-problems but also giving them access to
-
-um you know safely and securely to our
-
-provider Network to our uh you know to
-
-to our Outpatient Clinic Network which
-
-is the largest owned and operated
-
-Network in Canada by far uh so
-
-um and and when these and it's it was
-
-remarkable when we announced this
-
-program we've had just in the in the
-
-first uh week to 10 days we've had over
-
-a hundred uh inbound prospects come in
-
-uh that that wanted to you know
-
-collaborate with us and again I don't
-
-think that's necessarily for the money
-
-you know we're saying we would invest a
-
-minimum of a quarter of a million
-
-dollars you know a lot of them will
-
-likely be higher than a quarter of a
-
-million dollars
-
-so it's not life-changing money but but
-
-our structural advantages and and and
-
-the benefits that we have in the Well
-
-Network those are extremely hard to come
-
-by uh and I think and I think uh uh
-
-you'll see us uh you know help some of
-
-these companies uh succeed and they will
-
-help us drive uh you know more
-
-Innovation to that helps the provider
-
-but speaking of this very interesting AI
-
-I know your company just launched well
-
-AI voice this is super interesting tell
-
-me what it is and the impact it could
-
-have on health care providers
-
-yeah thanks for uh asking Richard our
-
-providers uh are thrilled with this you
-
-know we've we've had a number of of of
-
-our own well providers testing this
-
-technology and it it it really feels
-
-like magic to them it's essentially an
-
-ambient AI powered scribe so it's a it's
-
-a service that with the consent of the
-
-parties involved listens to the
-
-conversation between a patient and
-
-provider and then uh essentially
-
-condenses that into a medically relevant
-
-note for the chart files uh typically
-
-that is a lengthy process a doctor has
-
-to transcribe notes then review those
-
-notes and make sure that uh a a a a
-
-appropriate medically oriented and
-
-structured node is is is uh prepared and
-
-put into the chart and that could take
-
-you know sometimes more than more time
-
-than the actual consultation uh time and
-
-so we believe that on average if it's
-
-used regularly and consistently this can
-
-give providers back at least a third of
-
-their day
-
-um and and it's it's just a game changer
-
-uh and and uh we have now gone into
-
-General release with this product it's
-
-widely available in Canada uh it has
-
-been integrated into our EMR which makes
-
-it even more valuable tools like this
-
-are going to start popping up but if
-
-they're not integrated into your
-
-practice management system then you have
-
-to kind of have data in in more than one
-
-place and and move that around a little
-
-bit which which makes it a little bit
-
-more difficult especially with HIPAA
-
-requirements and and regulations so
-
-again I think this is the first of many
-
-types of different products and services
-
-that allow doctors to place more
-
-emphasis and focus on the patient
-
-experience instead of having their head
-
-in a laptop and looking at you once in a
-
-while they'll be looking at you and
-
-speaking to their practice management
-
-system and I think this you know think
-
-about it as Alexa for for our doctors uh
-
-you know this this ability to speak uh
-
-and and have you know uh you know Voice
-
-driven AI assistant that does things
-
-like this I think are going to be you
-
-know incredibly helpful and valuable uh
-
-for for healthcare providers
-
-super fascinating I mean we're just
-
-hearing you know more about AI maybe AI
-
-for the first time but here you are with
-
-a product already on the market in the
-
-in the healthcare field that's going to
-
-be pretty attractive to be out there uh
-
-right ahead of many other people right
-
-thank you Richard thanks for that
-
-recognition that's been Our intention we
-
-we want to demonstrate that we uh you
-
-know that we're all in on ensuring that
-
-technology that benefits providers uh is
-
-is is accelerated and uh de-risked and
-
-provided uh you know um in in a timely
-
-way you know providers need this help we
-
-we have a healthcare crisis in the
-
-country that is generally characterized
-
-as a as a lack of doctors and so imagine
-
-if we can get our doctors to be 20 or 30
-
-percent more productive through the use
-
-of these types of tools well they're
-
-going to just see more patience and and
-
-that's going to help all of us and uh
-
-and look if you step back Wells business
-
-model is all about having exposure to
-
-the success of doctors and doing our
-
-best to help them be more successful
-
-because we're in a revenue share
-
-relationship with most of the doctors
-
-that we work with and so this uh this is
-
-good for the ecosystem it's great for
-
-the provider and it's great for well as
-
-well super fascinating I'm Ed shabazzi
-
-CEO well Health Technologies ticker
-
-w-e-l-l great to catch up again thank
-
-you sir
-
-thank you Richard appreciate you having
-
-me
-
-[Music]
-
-thank you
-

server/evaluate/reference_texts/ref_sample_3.txt

@@ -1,970 +0,0 @@
-learning medicine is hard work osmosis
-
-makes it easy it takes our lectures and
-
-notes to create a personalized study
-
-plan with exclusive videos practice
-
-questions and flashcards and so much
-
-more try it free today
-
-in diabetes mellitus your body has
-
-trouble moving glucose which is the type
-
-of sugar from your blood into your cells
-
-this leads to high levels of glucose in
-
-your blood and not enough of it in your
-
-cells and remember that your cells need
-
-glucose as a source of energy so not
-
-letting the glucose enter means that the
-
-cells star for energy despite having
-
-glucose right on their doorstep in
-
-general the body controls how much
-
-glucose is in the blood relative to how
-
-much gets into the cells with two
-
-hormones insulin and glucagon insulin is
-
-used to reduce blood glucose levels and
-
-glucagon is used to increase blood
-
-glucose levels both of these hormones
-
-are produced by clusters of cells in the
-
-pancreas called islets of langerhans
-
-insulin is secreted by beta cells in the
-
-center of these islets and glucagon is
-
-secreted by alpha cells in the periphery
-
-of the islets insulin reduces the amount
-
-of glucose in the blood by binding to
-
-insulin receptors embedded in the cell
-
-membrane of various insulin responsive
-
-tissues like muscle cells in adipose
-
-tissue when activated the insulin
-
-receptors cause vesicles containing
-
-glucose transporter that are inside the
-
-cell to fuse with the cell membrane
-
-allowing glucose to be transported into
-
-the cell glucagon does exactly the
-
-opposite it raises the blood glucose
-
-levels by getting the liver to generate
-
-new molecules of glucose from other
-
-molecules and also break down glycogen
-
-into glucose so that I can all get
-
-dumped into the blood diabetes mellitus
-
-is diagnosed when blood glucose levels
-
-get too high and this is seen among 10
-
-percent of the world population there
-
-are two types of diabetes type 1 and
-
-type 2 and the main difference between
-
-them is the underlying mechanism that
-
-causes the blood glucose levels to rise
-
-about 10% of people with diabetes have
-
-type 1 and the remaining 90% of people
-
-with diabetes have type 2 let's start
-
-with type 1 diabetes mellitus sometimes
-
-just called type 1 diabetes in this
-
-situation the body doesn't make enough
-
-insulin the reason this happens is that
-
-in type 1 diabetes there's a type 4
-
-hypersensitivity response or a cell
-
-mediated immune response where a
-
-person's own T cells at
-
-the pancreas as a quick review remember
-
-that the immune system has T cells that
-
-react to all sorts of antigens which are
-
-usually small peptides polysaccharides
-
-or lipids and that some of these
-
-antigens are part of our own body cells
-
-it doesn't make sense to allow T cells
-
-that will attack our own cells to hang
-
-around until there's this process to
-
-eliminate them called self tolerance in
-
-type 1 diabetes there's a genetic
-
-abnormality that causes a loss of self
-
-tolerance among T cells that
-
-specifically target the beta cell
-
-antigens losing self tolerance means
-
-that these T cells are allowed to
-
-recruit other immune cells and
-
-coordinate an attack on these beta cells
-
-losing beta cells means less insulin and
-
-less insulin means that glucose piles up
-
-in the blood because it can't enter the
-
-body's cells one really important group
-
-of genes involved in regulation of the
-
-immune response is the human leukocyte
-
-antigen system or HLA system even though
-
-it's called a system it's basically this
-
-group of genes on chromosome 6 that
-
-encode the major histocompatibility
-
-complex or MHC which is a protein that's
-
-extremely important in helping the
-
-immune system recognize foreign
-
-molecules as well as maintaining self
-
-tolerance MHC is like the serving
-
-platter that antigens are presented to
-
-the immune cells on interestingly people
-
-with type 1 diabetes often have specific
-
-HLA genes in common with each other one
-
-called
-
-HLA dr3 and another called HLA dr4 but
-
-this is just a genetic clue right
-
-because not everyone with HLA dr3 and
-
-HLA dr4 develops diabetes in diabetes
-
-mellitus type 1 destruction of beta
-
-cells usually starts early in life but
-
-sometimes up to 90% of the beta cells
-
-are destroyed before symptoms crop up
-
-for clinical symptoms of uncontrolled
-
-diabetes that all sound similar our
-
-polyphagia glycosuria polyuria and
-
-polydipsia let's go through them one by
-
-one even though there's a lot of glucose
-
-in the blood it cannot get into the
-
-cells which leaves cells starved for
-
-energy so in response adipose tissue
-
-starts breaking down fat called
-
-lipolysis
-
-and muscle tissue starts breaking down
-
-proteins both of which results in weight
-
-loss for someone with uncontrolled
-
-diabetes this catabolic state leaves
-
-people feeling hungry
-
-also known as poly fascia Faiza means
-
-eating and poly means a lot now with
-
-high glucose levels that means that when
-
-blood gets filtered through the kidneys
-
-some of it starts to spill into the
-
-urine called glycosuria glyco surfers to
-
-glucose and urea the urine since glucose
-
-is osmotically active water tends to
-
-follow it resulting in an increase in
-
-urination or polyuria poly again refers
-
-to a lot and urea again refers to urine
-
-finally because there's so much
-
-urination people with uncontrolled
-
-diabetes become dehydrated and thirsty
-
-or polydipsia poly means a lot and dip
-
-SIA means thirst even though people with
-
-diabetes are not able to produce their
-
-own insulin they can still respond to
-
-insulin so treatment involves lifelong
-
-insulin therapy to regulate their blood
-
-glucose levels and basically enable
-
-their cells to use glucose
-
-one really serious complication with
-
-type 1 diabetes is called diabetic
-
-ketoacidosis or DKA to understand it
-
-let's go back to the process of
-
-lipolysis where fat is broken down into
-
-free fatty acids after that happens the
-
-liver turns the fatty acids into ketone
-
-bodies like Osito acetic acid in beta
-
-hydroxy butyrate acid a seed of acetic
-
-acid is a keto acid because it has a
-
-ketone group in a carboxylic acid group
-
-beta hydroxy rhetoric acid on the other
-
-hand even though it's still one of the
-
-ketone bodies isn't technically a keto
-
-acid since its ketone group has been
-
-reduced to a hydroxyl group these ketone
-
-bodies are important because they can be
-
-used by cells for energy but they also
-
-increase the acidity of the blood which
-
-is why it's called ketoacidosis and the
-
-blood becoming really acidic can have
-
-major effects throughout the body
-
-individuals can develop custom all
-
-respiration which is a deep and labored
-
-breathing as the body tries to move
-
-carbon dioxide out of the blood in an
-
-effort to reduce its acidity cells also
-
-have a transporter that exchanges
-
-hydrogen ions or protons for potassium
-
-when the blood gets acidic it's by
-
-definition loaded with protons that get
-
-sent into cells while potassium gets
-
-sent into the fluid outside cells
-
-another thing to keep in mind is that in
-
-addition to helping glucose enter cells
-
-insulin stimulates the sodium potassium
-
-ATPase --is which help potassium get
-
-into the cells and so without insulin
-
-more potassium stays in the fluid
-
-outside cells both of these mechanisms
-
-lead to increased potassium in the fluid
-
-outside cells which quickly makes it
-
-into the blood and causes hyperkalemia
-
-the potassium is then excreted so over
-
-time even though the blood potassium
-
-levels remain high over all stores of
-
-potassium in the body which include
-
-potassium inside cells starts to run low
-
-individuals will also have a high anion
-
-gap which reflects a large difference in
-
-the unmeasured negative and positive
-
-ions in the serum largely due to the
-
-build-up of ketoacids
-
-diabetic ketoacidosis can happen even in
-
-people who have already been diagnosed
-
-with diabetes and currently have some
-
-sort of insulin therapy
-
-in states of stress like an infection
-
-the body releases epinephrine which in
-
-turn stimulates the release of glucagon
-
-too much glucagon can tip the delicate
-
-hormonal balance of glucagon and insulin
-
-in favor of elevating blood sugars and
-
-can lead to a cascade of events we just
-
-described increased glucose in the blood
-
-loss of glucose in the urine loss of
-
-water dehydration and in parallel and
-
-need for alternative energy generation
-
-of ketone bodies and ketoacidosis
-
-interestingly both ketone bodies break
-
-down into acetone and escape as a gas by
-
-getting breathed out the lungs which
-
-gives us sweet fruity smell to a
-
-person's breath in general though that's
-
-the only sweet thing about this illness
-
-which also causes nausea vomiting and if
-
-severe mental status changes and acute
-
-cerebral edema
-
-treatment of a DKA episode involves
-
-giving plenty of fluids which helps with
-
-dehydration insulin which helps lower
-
-blood glucose levels and replacement of
-
-electrolytes like potassium all of which
-
-help to reverse the acidosis now let's
-
-switch gears and talk about type 2
-
-diabetes which is where the body makes
-
-insulin but the tissues don't respond as
-
-well to it the exact reason why cells
-
-don't respond isn't fully understood
-
-essentially the body's providing the
-
-normal amount of insulin but the cells
-
-don't move their glucose transporters to
-
-their membrane in response which
-
-remember is needed for the glucose to
-
-get into the cells these cells therefore
-
-have insulin resistance some risk
-
-factors for insulin resistance are
-
-obesity lack of exercise and
-
-hypertension the exact mechanisms are
-
-still being explored for example in
-
-excess of adipose tissue or fat is
-
-thought to cause the release of free
-
-fatty acids in so-called edible kinds
-
-which are signaling molecules that can
-
-cause inflammation which seems related
-
-to insulin resistance
-
-however many people that are obese are
-
-not diabetic so genetic factors probably
-
-play a major role as well we see this
-
-when we look at twin studies as well
-
-we're having a twin with type-2 diabetes
-
-increases the risk of developing type 2
-
-diabetes completely independently of
-
-other environmental risk factors in type
-
-2 diabetes since tissues don't respond
-
-as well to normal levels of insulin the
-
-body ends up producing more insulin in
-
-order to get the same effect and move
-
-glucose out of the blood
-
-they do this through beta cell
-
-hyperplasia an increased number of beta
-
-cells and beta cell hypertrophy where
-
-they actually grow in size all in this
-
-attempt to pump out more insulin this
-
-works for a while and by keeping insulin
-
-levels higher than normal blood glucose
-
-levels can be kept normal called normal
-
-glycemia now along with insulin beta
-
-cells also secrete islet amyloid
-
-polypeptide or amylin so while beta
-
-cells are cranking out insulin they also
-
-secrete an increased amount of amylin
-
-over time Emlyn builds up and aggregates
-
-in the islets this beta cell
-
-compensation though is not sustainable
-
-and over time those maxed out beta cells
-
-get exhausted and they become
-
-dysfunctional and undergo hypo trophy
-
-and get smaller as well as hypoplasia
-
-and die off as beta cells are lost in
-
-insulin levels decrease glucose levels
-
-in the blood start to increase in
-
-patients develop hyperglycemia which
-
-leads to similar clinical signs that we
-
-mentioned before like Paul aphasia
-
-glycosuria polyuria polydipsia but
-
-unlike type 1 diabetes there's generally
-
-some circulating insulin in type 2
-
-diabetes from the beta cells that are
-
-trying to compensate for the insulin
-
-resistance this means that the insulin
-
-glucagon balances such that diabetic
-
-ketoacidosis does not usually develop
-
-having said that a complication called
-
-hyperosmolar hyperglycemic state or HHS
-
-is much more common in type 2 diabetes
-
-than type 1 diabetes and it causes
-
-increased plasma osmolarity due to
-
-extreme dehydration and concentration of
-
-the blood to help understand this
-
-remember that glucose is a polar
-
-molecule that cannot passively diffuse
-
-across cell membranes which means that
-
-it acts as a solute so when levels of
-
-glucose are super high in the blood
-
-meaning it's a hyperosmolar State water
-
-starts to leave the body cells and enter
-
-the blood vessels leaving the cells were
-
-relatively dry in travailed rather than
-
-plump and juicy blood vessels that are
-
-full of water lead to increased
-
-urination and total body dehydration and
-
-this is a very serious situation because
-
-the dehydration of the body's cells and
-
-in particular the brain can cause a
-
-number of symptoms including mental
-
-status changes in HHS you can sometimes
-
-see mild ketone emia and acidosis but
-
-not to the extent that it's seen in DKA
-
-and in DKA you can see some hyper
-
-osmolarity so there's definitely overlap
-
-between these two syndromes
-
-besides type 1 and type 2 diabetes there
-
-are also a couple other subtypes of
-
-diabetes mellitus gestational diabetes
-
-is when pregnant women have increased
-
-blood glucose which is particularly
-
-during the third trimester although
-
-ultimately unknown the cause is thought
-
-to be related to pregnancy hormones that
-
-interfere with insulins action on
-
-insulin receptors also sometimes people
-
-can develop drug-induced diabetes which
-
-is where medications have side effects
-
-that tend to increase blood glucose
-
-levels the mechanism for both of these
-
-is thought to be related to insulin
-
-resistance like type 2 diabetes rather
-
-than an autoimmune destruction process
-
-like in type 1 diabetes diagnosing type
-
-1 or type 2 diabetes is done by getting
-
-a sense for how much glucose is floating
-
-around in the blood and has specific
-
-standards that the World Health
-
-Organization uses very commonly a
-
-fasting glucose test is taken where the
-
-person doesn't eat or drink except the
-
-water that's okay for a total of eight
-
-hours and then has their blood tested
-
-for glucose levels levels of 100
-
-milligrams per deciliter to 120
-
-five milligrams per deciliter indicates
-
-pre-diabetes and 126 milligrams per
-
-deciliter or higher indicates diabetes a
-
-non fasting a random glucose test can be
-
-done at any time with 200 milligrams per
-
-deciliter or higher being a red flag for
-
-diabetes another test is called an oral
-
-glucose tolerance test where person is
-
-given glucose and then blood samples are
-
-taken at time intervals to figure out
-
-how well it's being cleared from the
-
-blood the most important interval being
-
-two hours later levels of 140 milligrams
-
-per deciliter to 199 milligrams per
-
-deciliter indicate pre-diabetes
-
-and 200 or above indicates diabetes
-
-another thing to know is that when blood
-
-glucose levels get high the glucose can
-
-also stick to proteins that are floating
-
-around in the blood or in cells so that
-
-brings us to another type of test that
-
-can be done which is the hba1c test
-
-which tests for the proportion of
-
-hemoglobin in red blood cells that has
-
-glucose stuck to it called glycated
-
-hemoglobin hba1c levels of 5.7% 26.4%
-
-indicate pre-diabetes
-
-and 6.5 percent or higher indicates
-
-diabetes this proportion of glycated
-
-hemoglobin doesn't change day to day so
-
-it gives a sense for whether the blood
-
-glucose levels have been high over the
-
-past two to three months finally we have
-
-the c-peptide test which tests for
-
-byproducts of insulin production if the
-
-level of c-peptide is low or absent it
-
-means the pancreas is no longer
-
-producing enough insulin and the glucose
-
-cannot enter the cells
-
-for type one diabetes insulin is the
-
-only treatment option for type 2
-
-diabetes on the other hand lifestyle
-
-changes like weight loss and exercise
-
-along with a healthy diet and an oral
-
-anti-diabetic medication like metformin
-
-in several other classes can sometimes
-
-be enough to reverse some of that
-
-insulin resistance and keep blood sugar
-
-levels in check however if oral
-
-anti-diabetic medications fail type 2
-
-diabetes can also be treated with
-
-insulin something to bear in mind is
-
-that insulin treatment comes with a risk
-
-of hypoglycemia especially if insulin is
-
-taken without a meal symptoms of
-
-hypoglycemia can be mild like weakness
-
-hunger and shaking but they can progress
-
-to a loss of consciousness in seizures
-
-in severe cases in mild cases drinking
-
-juices or eating candy or sugar might be
-
-enough to bring blood sugar up but in
-
-severe cases intravenous glucose should
-
-be given as soon as possible
-
-the FDA has also recently approved
-
-intranasal glucagon as a treatment for
-
-severe hypoglycemia all right now over
-
-time high glucose levels can cause
-
-damage to tiny blood vessels while the
-
-micro vasculature in arterioles a
-
-process called hyaline
-
-arteriolosclerosis is where the walls of
-
-the arterioles develop hyaline deposits
-
-which are deposits of proteins and these
-
-make them hard and inflexible in
-
-capillaries the basement membrane can
-
-thicken and make it difficult for oxygen
-
-to easily move from the capillary to the
-
-tissues causing hypoxia
-
-one of the most significant effects is
-
-that diabetes increases the risk of
-
-medium and large arterial wall damage
-
-and subsequent atherosclerosis which can
-
-lead to heart attacks and strokes which
-
-are major causes of morbidity and
-
-mortality for patients with diabetes in
-
-the eyes diabetes can lead to
-
-retinopathy and evidence of that can be
-
-seen on a fundus copic exam that shows
-
-cotton-wool spots or flare hemorrhages
-
-and can eventually cause blindness in
-
-the kidneys the a ferrant and efferent
-
-arterioles as well as the glomerulus
-
-itself can get damaged which can lead to
-
-an F Radek syndrome that slowly
-
-diminishes the kidneys ability to filter
-
-blood over time and can ultimately lead
-
-to dialysis diabetes can also affect the
-
-function of nerves causing symptoms like
-
-a decrease in sensation in the toes and
-
-fingers sometimes called a stocking
-
-glove distribution as well as causes the
-
-autonomic nervous system to malfunction
-
-and that system controls a number of
-
-body functions
-
-everything from sweating to passing gas
-
-finally both the poor blood supply and
-
-nerve damage can lead to ulcers
-
-typically on the feet that don't heal
-
-quickly and can get pretty severe and
-
-need to be amputated these are some of
-
-the complications of uncontrolled
-
-diabetes which is why it's important to
-
-diagnose and control diabetes through a
-
-healthy lifestyle medications to reduce
-
-insulin resistance and even insulin
-
-therapy if beta cells have been
-
-exhausted while type 1 diabetes cannot
-
-be prevented type 2 diabetes can in fact
-
-many people with diabetes can control
-
-their blood sugar levels really
-
-effectively and live a full and active
-
-life without any of the complications
-
-thanks for watching if you're interested
-
-in a deeper dive on this topic take a
-
-look at as Moses org where we have
-
-flashcards questions and other awesome
-
-tools to help you learn medicine
-
-you
-

server/gpu/modal_deployments/README.md

@@ -1,81 +0,0 @@
-# Reflector GPU implementation - Transcription and LLM
-
-This repository hold an API for the GPU implementation of the Reflector API service,
-and use [Modal.com](https://modal.com)
-
-- `reflector_llm.py` - LLM API
-- `reflector_transcriber.py` - Transcription API
-
-## Modal.com deployment
-
-Create a modal secret, and name it `reflector-gpu`.
-It should contain an `REFLECTOR_APIKEY` environment variable with a value.
-
-The deployment is done using [Modal.com](https://modal.com) service.
-
-```
-$ modal deploy reflector_transcriber.py
-...
-└── 🔨 Created web => https://xxxx--reflector-transcriber-web.modal.run
-
-$ modal deploy reflector_llm.py
-...
-└── 🔨 Created web => https://xxxx--reflector-llm-web.modal.run
-```
-
-Then in your reflector api configuration `.env`, you can set theses keys:
-
-```
-TRANSCRIPT_BACKEND=modal
-TRANSCRIPT_URL=https://xxxx--reflector-transcriber-web.modal.run
-TRANSCRIPT_MODAL_API_KEY=REFLECTOR_APIKEY
-
-LLM_BACKEND=modal
-LLM_URL=https://xxxx--reflector-llm-web.modal.run
-LLM_MODAL_API_KEY=REFLECTOR_APIKEY
-```
-
-## API
-
-Authentication must be passed with the `Authorization` header, using the `bearer` scheme.
-
-```
-Authorization: bearer <REFLECTOR_APIKEY>
-```
-
-### LLM
-
-`POST /llm`
-
-**request**
-```
-{
-    "prompt": "xxx"
-}
-```
-
-**response**
-```
-{
-    "text": "xxx completed"
-}
-```
-
-### Transcription
-
-`POST /transcribe`
-
-**request** (multipart/form-data)
-
-- `file` - audio file
-- `language` - language code (e.g. `en`)
-
-**response**
-```
-{
-    "text": "xxx",
-    "words": [
-        {"text": "xxx", "start": 0.0, "end": 1.0}
-    ]
-}
-```

server/gpu/modal_deployments/reflector_diarizer.py

@@ -1,187 +0,0 @@
-"""
-Reflector GPU backend - diarizer
-===================================
-"""
-
-import os
-
-import modal.gpu
-from modal import App, Image, Secret, asgi_app, enter, method
-from pydantic import BaseModel
-
-PYANNOTE_MODEL_NAME: str = "pyannote/speaker-diarization-3.1"
-MODEL_DIR = "/root/diarization_models"
-app = App(name="reflector-diarizer")
-
-
-def migrate_cache_llm():
-    """
-    XXX The cache for model files in Transformers v4.22.0 has been updated.
-    Migrating your old cache. This is a one-time only operation. You can
-    interrupt this and resume the migration later on by calling
-    `transformers.utils.move_cache()`.
-    """
-    from transformers.utils.hub import move_cache
-
-    print("Moving LLM cache")
-    move_cache(cache_dir=MODEL_DIR, new_cache_dir=MODEL_DIR)
-    print("LLM cache moved")
-
-
-def download_pyannote_audio():
-    from pyannote.audio import Pipeline
-
-    Pipeline.from_pretrained(
-        PYANNOTE_MODEL_NAME,
-        cache_dir=MODEL_DIR,
-        use_auth_token=os.environ["HF_TOKEN"],
-    )
-
-
-diarizer_image = (
-    Image.debian_slim(python_version="3.10.8")
-    .pip_install(
-        "pyannote.audio==3.1.0",
-        "requests",
-        "onnx",
-        "torchaudio",
-        "onnxruntime-gpu",
-        "torch==2.0.0",
-        "transformers==4.34.0",
-        "sentencepiece",
-        "protobuf",
-        "numpy",
-        "huggingface_hub",
-        "hf-transfer",
-    )
-    .run_function(
-        download_pyannote_audio, secrets=[Secret.from_name("my-huggingface-secret")]
-    )
-    .run_function(migrate_cache_llm)
-    .env(
-        {
-            "LD_LIBRARY_PATH": (
-                "/usr/local/lib/python3.10/site-packages/nvidia/cudnn/lib/:"
-                "/opt/conda/lib/python3.10/site-packages/nvidia/cublas/lib/"
-            )
-        }
-    )
-)
-
-
-@app.cls(
-    gpu=modal.gpu.A100(size="40GB"),
-    timeout=60 * 30,
-    scaledown_window=60,
-    allow_concurrent_inputs=1,
-    image=diarizer_image,
-)
-class Diarizer:
-    @enter()
-    def enter(self):
-        import torch
-        from pyannote.audio import Pipeline
-
-        self.use_gpu = torch.cuda.is_available()
-        self.device = "cuda" if self.use_gpu else "cpu"
-        self.diarization_pipeline = Pipeline.from_pretrained(
-            PYANNOTE_MODEL_NAME, cache_dir=MODEL_DIR
-        )
-        self.diarization_pipeline.to(torch.device(self.device))
-
-    @method()
-    def diarize(self, audio_data: str, audio_suffix: str, timestamp: float):
-        import tempfile
-
-        import torchaudio
-
-        with tempfile.NamedTemporaryFile("wb+", suffix=f".{audio_suffix}") as fp:
-            fp.write(audio_data)
-
-            print("Diarizing audio")
-            waveform, sample_rate = torchaudio.load(fp.name)
-            diarization = self.diarization_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:]),
-                    }
-                )
-            print("Diarization complete")
-            return {"diarization": words}
-
-
-# -------------------------------------------------------------------
-# Web API
-# -------------------------------------------------------------------
-
-
-@app.function(
-    timeout=60 * 10,
-    scaledown_window=60 * 3,
-    allow_concurrent_inputs=40,
-    secrets=[
-        Secret.from_name("reflector-gpu"),
-    ],
-    image=diarizer_image,
-)
-@asgi_app()
-def web():
-    import requests
-    from fastapi import Depends, FastAPI, HTTPException, status
-    from fastapi.security import OAuth2PasswordBearer
-
-    diarizerstub = Diarizer()
-
-    app = FastAPI()
-
-    oauth2_scheme = OAuth2PasswordBearer(tokenUrl="token")
-
-    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"},
-            )
-
-    def validate_audio_file(audio_file_url: str):
-        # Check if the audio file exists
-        response = requests.head(audio_file_url, allow_redirects=True)
-        if response.status_code == 404:
-            raise HTTPException(
-                status_code=response.status_code,
-                detail="The audio file does not exist.",
-            )
-
-    class DiarizationResponse(BaseModel):
-        result: dict
-
-    @app.post(
-        "/diarize", dependencies=[Depends(apikey_auth), Depends(validate_audio_file)]
-    )
-    def diarize(
-        audio_file_url: str, timestamp: float = 0.0
-    ) -> HTTPException | DiarizationResponse:
-        # Currently the uploaded files are in mp3 format
-        audio_suffix = "mp3"
-
-        print("Downloading audio file")
-        response = requests.get(audio_file_url, allow_redirects=True)
-        print("Audio file downloaded successfully")
-
-        func = diarizerstub.diarize.spawn(
-            audio_data=response.content, audio_suffix=audio_suffix, timestamp=timestamp
-        )
-        result = func.get()
-        return result
-
-    return app

server/gpu/modal_deployments/reflector_llm.py

@@ -1,214 +0,0 @@
-"""
-Reflector GPU backend - LLM
-===========================
-
-"""
-
-import json
-import os
-import threading
-from typing import Optional
-
-import modal
-from modal import App, Image, Secret, asgi_app, enter, exit, method
-
-# LLM
-LLM_MODEL: str = "lmsys/vicuna-13b-v1.5"
-LLM_LOW_CPU_MEM_USAGE: bool = True
-LLM_TORCH_DTYPE: str = "bfloat16"
-LLM_MAX_NEW_TOKENS: int = 300
-
-IMAGE_MODEL_DIR = "/root/llm_models"
-
-app = App(name="reflector-llm")
-
-
-def download_llm():
-    from huggingface_hub import snapshot_download
-
-    print("Downloading LLM model")
-    snapshot_download(LLM_MODEL, cache_dir=IMAGE_MODEL_DIR)
-    print("LLM model downloaded")
-
-
-def migrate_cache_llm():
-    """
-    XXX The cache for model files in Transformers v4.22.0 has been updated.
-    Migrating your old cache. This is a one-time only operation. You can
-    interrupt this and resume the migration later on by calling
-    `transformers.utils.move_cache()`.
-    """
-    from transformers.utils.hub import move_cache
-
-    print("Moving LLM cache")
-    move_cache(cache_dir=IMAGE_MODEL_DIR, new_cache_dir=IMAGE_MODEL_DIR)
-    print("LLM cache moved")
-
-
-llm_image = (
-    Image.debian_slim(python_version="3.10.8")
-    .apt_install("git")
-    .pip_install(
-        "transformers",
-        "torch",
-        "sentencepiece",
-        "protobuf",
-        "jsonformer==0.12.0",
-        "accelerate==0.21.0",
-        "einops==0.6.1",
-        "hf-transfer~=0.1",
-        "huggingface_hub==0.16.4",
-    )
-    .env({"HF_HUB_ENABLE_HF_TRANSFER": "1"})
-    .run_function(download_llm)
-    .run_function(migrate_cache_llm)
-)
-
-
-@app.cls(
-    gpu="A100",
-    timeout=60 * 5,
-    scaledown_window=60 * 5,
-    allow_concurrent_inputs=15,
-    image=llm_image,
-)
-class LLM:
-    @enter()
-    def enter(self):
-        import torch
-        from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig
-
-        print("Instance llm model")
-        model = AutoModelForCausalLM.from_pretrained(
-            LLM_MODEL,
-            torch_dtype=getattr(torch, LLM_TORCH_DTYPE),
-            low_cpu_mem_usage=LLM_LOW_CPU_MEM_USAGE,
-            cache_dir=IMAGE_MODEL_DIR,
-            local_files_only=True,
-        )
-
-        # JSONFormer doesn't yet support generation configs
-        print("Instance llm generation config")
-        model.config.max_new_tokens = LLM_MAX_NEW_TOKENS
-
-        # generation configuration
-        gen_cfg = GenerationConfig.from_model_config(model.config)
-        gen_cfg.max_new_tokens = LLM_MAX_NEW_TOKENS
-
-        # load tokenizer
-        print("Instance llm tokenizer")
-        tokenizer = AutoTokenizer.from_pretrained(
-            LLM_MODEL, cache_dir=IMAGE_MODEL_DIR, local_files_only=True
-        )
-
-        # move model to gpu
-        print("Move llm model to GPU")
-        model = model.cuda()
-
-        print("Warmup llm done")
-        self.model = model
-        self.tokenizer = tokenizer
-        self.gen_cfg = gen_cfg
-        self.GenerationConfig = GenerationConfig
-
-        self.lock = threading.Lock()
-
-    @exit()
-    def exit():
-        print("Exit llm")
-
-    @method()
-    def generate(
-        self, prompt: str, gen_schema: str | None, gen_cfg: str | None
-    ) -> dict:
-        """
-        Perform a generation action using the LLM
-        """
-        print(f"Generate {prompt=}")
-        if gen_cfg:
-            gen_cfg = self.GenerationConfig.from_dict(json.loads(gen_cfg))
-        else:
-            gen_cfg = self.gen_cfg
-
-        # If a gen_schema is given, conform to gen_schema
-        with self.lock:
-            if gen_schema:
-                import jsonformer
-
-                print(f"Schema {gen_schema=}")
-                jsonformer_llm = jsonformer.Jsonformer(
-                    model=self.model,
-                    tokenizer=self.tokenizer,
-                    json_schema=json.loads(gen_schema),
-                    prompt=prompt,
-                    max_string_token_length=gen_cfg.max_new_tokens,
-                )
-                response = jsonformer_llm()
-            else:
-                # If no gen_schema, perform prompt only generation
-
-                # tokenize prompt
-                input_ids = self.tokenizer.encode(prompt, return_tensors="pt").to(
-                    self.model.device
-                )
-                output = self.model.generate(input_ids, generation_config=gen_cfg)
-
-                # decode output
-                response = self.tokenizer.decode(
-                    output[0].cpu(), skip_special_tokens=True
-                )
-                response = response[len(prompt) :]
-        print(f"Generated {response=}")
-        return {"text": response}
-
-
-# -------------------------------------------------------------------
-# Web API
-# -------------------------------------------------------------------
-
-
-@app.function(
-    scaledown_window=60 * 10,
-    timeout=60 * 5,
-    allow_concurrent_inputs=45,
-    secrets=[
-        Secret.from_name("reflector-gpu"),
-    ],
-)
-@asgi_app()
-def web():
-    from fastapi import Depends, FastAPI, HTTPException, status
-    from fastapi.security import OAuth2PasswordBearer
-    from pydantic import BaseModel
-
-    llmstub = LLM()
-
-    app = FastAPI()
-    oauth2_scheme = OAuth2PasswordBearer(tokenUrl="token")
-
-    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 LLMRequest(BaseModel):
-        prompt: str
-        gen_schema: Optional[dict] = None
-        gen_cfg: Optional[dict] = None
-
-    @app.post("/llm", dependencies=[Depends(apikey_auth)])
-    def llm(
-        req: LLMRequest,
-    ):
-        gen_schema = json.dumps(req.gen_schema) if req.gen_schema else None
-        gen_cfg = json.dumps(req.gen_cfg) if req.gen_cfg else None
-        func = llmstub.generate.spawn(
-            prompt=req.prompt, gen_schema=gen_schema, gen_cfg=gen_cfg
-        )
-        result = func.get()
-        return result
-
-    return app

server/gpu/modal_deployments/reflector_llm_zephyr.py

@@ -1,220 +0,0 @@
-"""
-Reflector GPU backend - LLM
-===========================
-
-"""
-
-import json
-import os
-import threading
-from typing import Optional
-
-import modal
-from modal import App, Image, Secret, asgi_app, enter, exit, method
-
-# LLM
-LLM_MODEL: str = "HuggingFaceH4/zephyr-7b-alpha"
-LLM_LOW_CPU_MEM_USAGE: bool = True
-LLM_TORCH_DTYPE: str = "bfloat16"
-LLM_MAX_NEW_TOKENS: int = 300
-
-IMAGE_MODEL_DIR = "/root/llm_models/zephyr"
-
-app = App(name="reflector-llm-zephyr")
-
-
-def download_llm():
-    from huggingface_hub import snapshot_download
-
-    print("Downloading LLM model")
-    snapshot_download(LLM_MODEL, cache_dir=IMAGE_MODEL_DIR)
-    print("LLM model downloaded")
-
-
-def migrate_cache_llm():
-    """
-    XXX The cache for model files in Transformers v4.22.0 has been updated.
-    Migrating your old cache. This is a one-time only operation. You can
-    interrupt this and resume the migration later on by calling
-    `transformers.utils.move_cache()`.
-    """
-    from transformers.utils.hub import move_cache
-
-    print("Moving LLM cache")
-    move_cache(cache_dir=IMAGE_MODEL_DIR, new_cache_dir=IMAGE_MODEL_DIR)
-    print("LLM cache moved")
-
-
-llm_image = (
-    Image.debian_slim(python_version="3.10.8")
-    .apt_install("git")
-    .pip_install(
-        "transformers==4.34.0",
-        "torch",
-        "sentencepiece",
-        "protobuf",
-        "jsonformer==0.12.0",
-        "accelerate==0.21.0",
-        "einops==0.6.1",
-        "hf-transfer~=0.1",
-        "huggingface_hub==0.16.4",
-    )
-    .env({"HF_HUB_ENABLE_HF_TRANSFER": "1"})
-    .run_function(download_llm)
-    .run_function(migrate_cache_llm)
-)
-
-
-@app.cls(
-    gpu="A10G",
-    timeout=60 * 5,
-    scaledown_window=60 * 5,
-    allow_concurrent_inputs=10,
-    image=llm_image,
-)
-class LLM:
-    @enter()
-    def enter(self):
-        import torch
-        from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig
-
-        print("Instance llm model")
-        model = AutoModelForCausalLM.from_pretrained(
-            LLM_MODEL,
-            torch_dtype=getattr(torch, LLM_TORCH_DTYPE),
-            low_cpu_mem_usage=LLM_LOW_CPU_MEM_USAGE,
-            cache_dir=IMAGE_MODEL_DIR,
-            local_files_only=True,
-        )
-
-        # JSONFormer doesn't yet support generation configs
-        print("Instance llm generation config")
-        model.config.max_new_tokens = LLM_MAX_NEW_TOKENS
-
-        # generation configuration
-        gen_cfg = GenerationConfig.from_model_config(model.config)
-        gen_cfg.max_new_tokens = LLM_MAX_NEW_TOKENS
-
-        # load tokenizer
-        print("Instance llm tokenizer")
-        tokenizer = AutoTokenizer.from_pretrained(
-            LLM_MODEL, cache_dir=IMAGE_MODEL_DIR, local_files_only=True
-        )
-        gen_cfg.pad_token_id = tokenizer.eos_token_id
-        gen_cfg.eos_token_id = tokenizer.eos_token_id
-        tokenizer.pad_token = tokenizer.eos_token
-        model.config.pad_token_id = tokenizer.eos_token_id
-
-        # move model to gpu
-        print("Move llm model to GPU")
-        model = model.cuda()
-
-        print("Warmup llm done")
-        self.model = model
-        self.tokenizer = tokenizer
-        self.gen_cfg = gen_cfg
-        self.GenerationConfig = GenerationConfig
-        self.lock = threading.Lock()
-
-    @exit()
-    def exit():
-        print("Exit llm")
-
-    @method()
-    def generate(
-        self, prompt: str, gen_schema: str | None, gen_cfg: str | None
-    ) -> dict:
-        """
-        Perform a generation action using the LLM
-        """
-        print(f"Generate {prompt=}")
-        if gen_cfg:
-            gen_cfg = self.GenerationConfig.from_dict(json.loads(gen_cfg))
-            gen_cfg.pad_token_id = self.tokenizer.eos_token_id
-            gen_cfg.eos_token_id = self.tokenizer.eos_token_id
-        else:
-            gen_cfg = self.gen_cfg
-
-        # If a gen_schema is given, conform to gen_schema
-        with self.lock:
-            if gen_schema:
-                import jsonformer
-
-                print(f"Schema {gen_schema=}")
-                jsonformer_llm = jsonformer.Jsonformer(
-                    model=self.model,
-                    tokenizer=self.tokenizer,
-                    json_schema=json.loads(gen_schema),
-                    prompt=prompt,
-                    max_string_token_length=gen_cfg.max_new_tokens,
-                )
-                response = jsonformer_llm()
-            else:
-                # If no gen_schema, perform prompt only generation
-
-                # tokenize prompt
-                input_ids = self.tokenizer.encode(prompt, return_tensors="pt").to(
-                    self.model.device
-                )
-                output = self.model.generate(input_ids, generation_config=gen_cfg)
-
-                # decode output
-                response = self.tokenizer.decode(
-                    output[0].cpu(), skip_special_tokens=True
-                )
-                response = response[len(prompt) :]
-                response = {"long_summary": response}
-        print(f"Generated {response=}")
-        return {"text": response}
-
-
-# -------------------------------------------------------------------
-# Web API
-# -------------------------------------------------------------------
-
-
-@app.function(
-    scaledown_window=60 * 10,
-    timeout=60 * 5,
-    allow_concurrent_inputs=30,
-    secrets=[
-        Secret.from_name("reflector-gpu"),
-    ],
-)
-@asgi_app()
-def web():
-    from fastapi import Depends, FastAPI, HTTPException, status
-    from fastapi.security import OAuth2PasswordBearer
-    from pydantic import BaseModel
-
-    llmstub = LLM()
-
-    app = FastAPI()
-    oauth2_scheme = OAuth2PasswordBearer(tokenUrl="token")
-
-    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 LLMRequest(BaseModel):
-        prompt: str
-        gen_schema: Optional[dict] = None
-        gen_cfg: Optional[dict] = None
-
-    @app.post("/llm", dependencies=[Depends(apikey_auth)])
-    def llm(
-        req: LLMRequest,
-    ):
-        gen_schema = json.dumps(req.gen_schema) if req.gen_schema else None
-        gen_cfg = json.dumps(req.gen_cfg) if req.gen_cfg else None
-        func = llmstub.generate.spawn(
-            prompt=req.prompt, gen_schema=gen_schema, gen_cfg=gen_cfg
-        )
-        result = func.get()
-        return result
-
-    return app

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_vllm_hermes3.py

@@ -1,171 +0,0 @@
-# # Run an OpenAI-Compatible vLLM Server
-
-import modal
-
-MODELS_DIR = "/llamas"
-MODEL_NAME = "NousResearch/Hermes-3-Llama-3.1-8B"
-N_GPU = 1
-
-
-def download_llm():
-    from huggingface_hub import snapshot_download
-
-    print("Downloading LLM model")
-    snapshot_download(
-        MODEL_NAME,
-        local_dir=f"{MODELS_DIR}/{MODEL_NAME}",
-        ignore_patterns=[
-            "*.pt",
-            "*.bin",
-            "*.pth",
-            "original/*",
-        ],  # Ensure safetensors
-    )
-    print("LLM model downloaded")
-
-
-def move_cache():
-    from transformers.utils import move_cache as transformers_move_cache
-
-    transformers_move_cache()
-
-
-vllm_image = (
-    modal.Image.debian_slim(python_version="3.10")
-    .pip_install("vllm==0.5.3post1")
-    .env({"HF_HUB_ENABLE_HF_TRANSFER": "1"})
-    .pip_install(
-        # "accelerate==0.34.2",
-        "einops==0.8.0",
-        "hf-transfer~=0.1",
-    )
-    .run_function(download_llm)
-    .run_function(move_cache)
-    .pip_install(
-        "bitsandbytes>=0.42.9",
-    )
-)
-
-app = modal.App("reflector-vllm-hermes3")
-
-
-@app.function(
-    image=vllm_image,
-    gpu=modal.gpu.A100(count=N_GPU, size="40GB"),
-    timeout=60 * 5,
-    scaledown_window=60 * 5,
-    allow_concurrent_inputs=100,
-    secrets=[
-        modal.Secret.from_name("reflector-gpu"),
-    ],
-)
-@modal.asgi_app()
-def serve():
-    import os
-
-    import fastapi
-    import vllm.entrypoints.openai.api_server as api_server
-    from vllm.engine.arg_utils import AsyncEngineArgs
-    from vllm.engine.async_llm_engine import AsyncLLMEngine
-    from vllm.entrypoints.logger import RequestLogger
-    from vllm.entrypoints.openai.serving_chat import OpenAIServingChat
-    from vllm.entrypoints.openai.serving_completion import OpenAIServingCompletion
-    from vllm.usage.usage_lib import UsageContext
-
-    TOKEN = os.environ["REFLECTOR_GPU_APIKEY"]
-
-    # create a fastAPI app that uses vLLM's OpenAI-compatible router
-    web_app = fastapi.FastAPI(
-        title=f"OpenAI-compatible {MODEL_NAME} server",
-        description="Run an OpenAI-compatible LLM server with vLLM on modal.com",
-        version="0.0.1",
-        docs_url="/docs",
-    )
-
-    # security: CORS middleware for external requests
-    http_bearer = fastapi.security.HTTPBearer(
-        scheme_name="Bearer Token",
-        description="See code for authentication details.",
-    )
-    web_app.add_middleware(
-        fastapi.middleware.cors.CORSMiddleware,
-        allow_origins=["*"],
-        allow_credentials=True,
-        allow_methods=["*"],
-        allow_headers=["*"],
-    )
-
-    # security: inject dependency on authed routes
-    async def is_authenticated(api_key: str = fastapi.Security(http_bearer)):
-        if api_key.credentials != TOKEN:
-            raise fastapi.HTTPException(
-                status_code=fastapi.status.HTTP_401_UNAUTHORIZED,
-                detail="Invalid authentication credentials",
-            )
-        return {"username": "authenticated_user"}
-
-    router = fastapi.APIRouter(dependencies=[fastapi.Depends(is_authenticated)])
-
-    # wrap vllm's router in auth router
-    router.include_router(api_server.router)
-    # add authed vllm to our fastAPI app
-    web_app.include_router(router)
-
-    engine_args = AsyncEngineArgs(
-        model=MODELS_DIR + "/" + MODEL_NAME,
-        tensor_parallel_size=N_GPU,
-        gpu_memory_utilization=0.90,
-        # max_model_len=8096,
-        enforce_eager=False,  # capture the graph for faster inference, but slower cold starts (30s > 20s)
-        # --- 4 bits load
-        # quantization="bitsandbytes",
-        # load_format="bitsandbytes",
-    )
-
-    engine = AsyncLLMEngine.from_engine_args(
-        engine_args, usage_context=UsageContext.OPENAI_API_SERVER
-    )
-
-    model_config = get_model_config(engine)
-
-    request_logger = RequestLogger(max_log_len=2048)
-
-    api_server.openai_serving_chat = OpenAIServingChat(
-        engine,
-        model_config=model_config,
-        served_model_names=[MODEL_NAME],
-        chat_template=None,
-        response_role="assistant",
-        lora_modules=[],
-        prompt_adapters=[],
-        request_logger=request_logger,
-    )
-    api_server.openai_serving_completion = OpenAIServingCompletion(
-        engine,
-        model_config=model_config,
-        served_model_names=[MODEL_NAME],
-        lora_modules=[],
-        prompt_adapters=[],
-        request_logger=request_logger,
-    )
-
-    return web_app
-
-
-def get_model_config(engine):
-    import asyncio
-
-    try:  # adapted from vLLM source -- https://github.com/vllm-project/vllm/blob/507ef787d85dec24490069ffceacbd6b161f4f72/vllm/entrypoints/openai/api_server.py#L235C1-L247C1
-        event_loop = asyncio.get_running_loop()
-    except RuntimeError:
-        event_loop = None
-
-    if event_loop is not None and event_loop.is_running():
-        # If the current is instanced by Ray Serve,
-        # there is already a running event loop
-        model_config = event_loop.run_until_complete(engine.get_model_config())
-    else:
-        # When using single vLLM without engine_use_ray
-        model_config = asyncio.run(engine.get_model_config())
-
-    return model_config

server/migrations/README

@@ -1 +1,3 @@
-Generic single-database configuration.
+Generic single-database configuration.
+
+Both data migrations and schema migrations must be in migrations.

server/migrations/env.py

@@ -1,9 +1,10 @@
 from logging.config import fileConfig
 
 from alembic import context
+from sqlalchemy import engine_from_config, pool
+
 from reflector.db import metadata
 from reflector.settings import settings
-from sqlalchemy import engine_from_config, pool
 
 # this is the Alembic Config object, which provides
 # access to the values within the .ini file in use.

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/0925da921477_unique_room_names.py

@@ -8,7 +8,6 @@ Create Date: 2024-09-24 16:12:56.944133
 
 from typing import Sequence, Union
 
-import sqlalchemy as sa
 from alembic import op
 
 # revision identifiers, used by Alembic.

server/migrations/versions/0ab2d7ffaa16_add_long_summary_to_search_vector.py

@@ -0,0 +1,64 @@
+"""add_long_summary_to_search_vector
+
+Revision ID: 0ab2d7ffaa16
+Revises: b1c33bd09963
+Create Date: 2025-08-15 13:27:52.680211
+
+"""
+
+from typing import Sequence, Union
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "0ab2d7ffaa16"
+down_revision: Union[str, None] = "b1c33bd09963"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # Drop the existing search vector column and index
+    op.drop_index("idx_transcript_search_vector_en", table_name="transcript")
+    op.drop_column("transcript", "search_vector_en")
+
+    # Recreate the search vector column with long_summary included
+    op.execute("""
+        ALTER TABLE transcript ADD COLUMN search_vector_en tsvector
+        GENERATED ALWAYS AS (
+            setweight(to_tsvector('english', coalesce(title, '')), 'A') ||
+            setweight(to_tsvector('english', coalesce(long_summary, '')), 'B') ||
+            setweight(to_tsvector('english', coalesce(webvtt, '')), 'C')
+        ) STORED
+    """)
+
+    # Recreate the GIN index for the search vector
+    op.create_index(
+        "idx_transcript_search_vector_en",
+        "transcript",
+        ["search_vector_en"],
+        postgresql_using="gin",
+    )
+
+
+def downgrade() -> None:
+    # Drop the updated search vector column and index
+    op.drop_index("idx_transcript_search_vector_en", table_name="transcript")
+    op.drop_column("transcript", "search_vector_en")
+
+    # Recreate the original search vector column without long_summary
+    op.execute("""
+        ALTER TABLE transcript ADD COLUMN search_vector_en tsvector
+        GENERATED ALWAYS AS (
+            setweight(to_tsvector('english', coalesce(title, '')), 'A') ||
+            setweight(to_tsvector('english', coalesce(webvtt, '')), 'B')
+        ) STORED
+    """)
+
+    # Recreate the GIN index for the search vector
+    op.create_index(
+        "idx_transcript_search_vector_en",
+        "transcript",
+        ["search_vector_en"],
+        postgresql_using="gin",
+    )

server/migrations/versions/0bc0f3ff0111_add_webvtt_field_to_transcript.py

@@ -0,0 +1,25 @@
+"""add_webvtt_field_to_transcript
+
+Revision ID: 0bc0f3ff0111
+Revises: b7df9609542c
+Create Date: 2025-08-05 19:36:41.740957
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "0bc0f3ff0111"
+down_revision: Union[str, None] = "b7df9609542c"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.add_column("transcript", sa.Column("webvtt", sa.Text(), nullable=True))
+
+
+def downgrade() -> None:
+    op.drop_column("transcript", "webvtt")

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/0fea6d96b096_add_share_mode.py

@@ -5,11 +5,11 @@ Revises: f819277e5169
 Create Date: 2023-11-07 11:12:21.614198
 
 """
+
 from typing import Sequence, Union
 
-from alembic import op
 import sqlalchemy as sa
-
+from alembic import op
 
 # revision identifiers, used by Alembic.
 revision: str = "0fea6d96b096"

server/migrations/versions/116b2f287eab_add_full_text_search.py

@@ -0,0 +1,46 @@
+"""add_full_text_search
+
+Revision ID: 116b2f287eab
+Revises: 0bc0f3ff0111
+Create Date: 2025-08-07 11:27:38.473517
+
+"""
+
+from typing import Sequence, Union
+
+from alembic import op
+
+revision: str = "116b2f287eab"
+down_revision: Union[str, None] = "0bc0f3ff0111"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    conn = op.get_bind()
+    if conn.dialect.name != "postgresql":
+        return
+
+    op.execute("""
+        ALTER TABLE transcript ADD COLUMN search_vector_en tsvector
+        GENERATED ALWAYS AS (
+            setweight(to_tsvector('english', coalesce(title, '')), 'A') ||
+            setweight(to_tsvector('english', coalesce(webvtt, '')), 'B')
+        ) STORED
+    """)
+
+    op.create_index(
+        "idx_transcript_search_vector_en",
+        "transcript",
+        ["search_vector_en"],
+        postgresql_using="gin",
+    )
+
+
+def downgrade() -> None:
+    conn = op.get_bind()
+    if conn.dialect.name != "postgresql":
+        return
+
+    op.drop_index("idx_transcript_search_vector_en", table_name="transcript")
+    op.drop_column("transcript", "search_vector_en")

server/migrations/versions/125031f7cb78_participants.py

@@ -5,26 +5,26 @@ Revises: 0fea6d96b096
 Create Date: 2023-11-30 15:56:03.341466
 
 """
+
 from typing import Sequence, Union
 
-from alembic import op
 import sqlalchemy as sa
-
+from alembic import op
 
 # revision identifiers, used by Alembic.
-revision: str = '125031f7cb78'
-down_revision: Union[str, None] = '0fea6d96b096'
+revision: str = "125031f7cb78"
+down_revision: Union[str, None] = "0fea6d96b096"
 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.add_column('transcript', sa.Column('participants', sa.JSON(), nullable=True))
+    op.add_column("transcript", sa.Column("participants", sa.JSON(), nullable=True))
     # ### end Alembic commands ###
 
 
 def downgrade() -> None:
     # ### commands auto generated by Alembic - please adjust! ###
-    op.drop_column('transcript', 'participants')
+    op.drop_column("transcript", "participants")
     # ### 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/20250617140003_add_meeting_consent_table.py

@@ -5,6 +5,7 @@ Revises: f819277e5169
 Create Date: 2025-06-17 14:00:03.000000
 
 """
+
 from typing import Sequence, Union
 
 import sqlalchemy as sa
@@ -19,16 +20,16 @@ depends_on: Union[str, Sequence[str], None] = None
 
 def upgrade() -> None:
     op.create_table(
-        'meeting_consent',
-        sa.Column('id', sa.String(), nullable=False),
-        sa.Column('meeting_id', sa.String(), nullable=False),
-        sa.Column('user_id', sa.String(), nullable=True),
-        sa.Column('consent_given', sa.Boolean(), nullable=False),
-        sa.Column('consent_timestamp', sa.DateTime(), nullable=False),
-        sa.PrimaryKeyConstraint('id'),
-        sa.ForeignKeyConstraint(['meeting_id'], ['meeting.id']),
+        "meeting_consent",
+        sa.Column("id", sa.String(), nullable=False),
+        sa.Column("meeting_id", sa.String(), nullable=False),
+        sa.Column("user_id", sa.String(), nullable=True),
+        sa.Column("consent_given", sa.Boolean(), nullable=False),
+        sa.Column("consent_timestamp", sa.DateTime(), nullable=False),
+        sa.PrimaryKeyConstraint("id"),
+        sa.ForeignKeyConstraint(["meeting_id"], ["meeting.id"]),
     )
 
 
 def downgrade() -> None:
-    op.drop_table('meeting_consent')
+    op.drop_table("meeting_consent")

server/migrations/versions/20250618140000_add_audio_deleted_field_to_transcript.py

@@ -5,6 +5,7 @@ Revises: 20250617140003
 Create Date: 2025-06-18 14:00:00.000000
 
 """
+
 from typing import Sequence, Union
 
 import sqlalchemy as sa
@@ -22,4 +23,4 @@ def upgrade() -> None:
 
 
 def downgrade() -> None:
-    op.drop_column("transcript", "audio_deleted")
+    op.drop_column("transcript", "audio_deleted")

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/2cf0b60a9d34_fix_transcript_duration_type.py

@@ -5,36 +5,40 @@ Revises: ccd68dc784ff
 Create Date: 2025-07-15 16:53:40.397394
 
 """
+
 from typing import Sequence, Union
 
-from alembic import op
 import sqlalchemy as sa
-
+from alembic import op
 
 # revision identifiers, used by Alembic.
-revision: str = '2cf0b60a9d34'
-down_revision: Union[str, None] = 'ccd68dc784ff'
+revision: str = "2cf0b60a9d34"
+down_revision: Union[str, None] = "ccd68dc784ff"
 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('transcript', schema=None) as batch_op:
-        batch_op.alter_column('duration',
-               existing_type=sa.INTEGER(),
-               type_=sa.Float(),
-               existing_nullable=True)
+    with op.batch_alter_table("transcript", schema=None) as batch_op:
+        batch_op.alter_column(
+            "duration",
+            existing_type=sa.INTEGER(),
+            type_=sa.Float(),
+            existing_nullable=True,
+        )
 
     # ### end Alembic commands ###
 
 
 def downgrade() -> None:
     # ### commands auto generated by Alembic - please adjust! ###
-    with op.batch_alter_table('transcript', schema=None) as batch_op:
-        batch_op.alter_column('duration',
-               existing_type=sa.Float(),
-               type_=sa.INTEGER(),
-               existing_nullable=True)
+    with op.batch_alter_table("transcript", schema=None) as batch_op:
+        batch_op.alter_column(
+            "duration",
+            existing_type=sa.Float(),
+            type_=sa.INTEGER(),
+            existing_nullable=True,
+        )
 
     # ### end Alembic commands ###

server/migrations/versions/38a927dcb099_rename_back_text_to_transcript.py

@@ -5,17 +5,17 @@ Revises: 9920ecfe2735
 Create Date: 2023-11-02 19:53:09.116240
 
 """
+
 from typing import Sequence, Union
 
-from alembic import op
 import sqlalchemy as sa
-from sqlalchemy.sql import table, column
+from alembic import op
 from sqlalchemy import select
-
+from sqlalchemy.sql import column, table
 
 # revision identifiers, used by Alembic.
-revision: str = '38a927dcb099'
-down_revision: Union[str, None] = '9920ecfe2735'
+revision: str = "38a927dcb099"
+down_revision: Union[str, None] = "9920ecfe2735"
 branch_labels: Union[str, Sequence[str], None] = None
 depends_on: Union[str, Sequence[str], None] = None
 

server/migrations/versions/4814901632bc_fix_duration.py

@@ -5,13 +5,13 @@ Revises: 38a927dcb099
 Create Date: 2023-11-10 18:12:17.886522
 
 """
+
 from typing import Sequence, Union
 
-from alembic import op
 import sqlalchemy as sa
-from sqlalchemy.sql import table, column
+from alembic import op
 from sqlalchemy import select
-
+from sqlalchemy.sql import column, table
 
 # revision identifiers, used by Alembic.
 revision: str = "4814901632bc"
@@ -24,9 +24,11 @@ def upgrade() -> None:
     # for all the transcripts, calculate the duration from the mp3
     # and update the duration column
     from pathlib import Path
-    from reflector.settings import settings
+
     import av
 
+    from reflector.settings import settings
+
     bind = op.get_bind()
     transcript = table(
         "transcript", column("id", sa.String), column("duration", sa.Float)

server/migrations/versions/543ed284d69a_init.py

@@ -5,14 +5,11 @@ Revises:
 Create Date: 2023-08-29 10:54:45.142974
 
 """
+
 from typing import Sequence, Union
 
-from alembic import op
-import sqlalchemy as sa
-
-
 # revision identifiers, used by Alembic.
-revision: str = '543ed284d69a'
+revision: str = "543ed284d69a"
 down_revision: Union[str, None] = None
 branch_labels: Union[str, Sequence[str], None] = None
 depends_on: Union[str, Sequence[str], None] = None

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/62dea3db63a5_add_room_options.py

@@ -8,9 +8,8 @@ Create Date: 2025-06-27 09:04:21.006823
 
 from typing import Sequence, Union
 
-from alembic import op
 import sqlalchemy as sa
-
+from alembic import op
 
 # revision identifiers, used by Alembic.
 revision: str = "62dea3db63a5"
@@ -33,7 +32,7 @@ def upgrade() -> None:
         sa.Column("user_id", sa.String(), nullable=True),
         sa.Column("room_id", sa.String(), nullable=True),
         sa.Column(
-            "is_locked", sa.Boolean(), server_default=sa.text("0"), nullable=False
+            "is_locked", sa.Boolean(), server_default=sa.text("false"), nullable=False
         ),
         sa.Column("room_mode", sa.String(), server_default="normal", nullable=False),
         sa.Column(
@@ -54,12 +53,15 @@ def upgrade() -> None:
         sa.Column("user_id", sa.String(), nullable=False),
         sa.Column("created_at", sa.DateTime(), nullable=False),
         sa.Column(
-            "zulip_auto_post", sa.Boolean(), server_default=sa.text("0"), nullable=False
+            "zulip_auto_post",
+            sa.Boolean(),
+            server_default=sa.text("false"),
+            nullable=False,
         ),
         sa.Column("zulip_stream", sa.String(), nullable=True),
         sa.Column("zulip_topic", sa.String(), nullable=True),
         sa.Column(
-            "is_locked", sa.Boolean(), server_default=sa.text("0"), nullable=False
+            "is_locked", sa.Boolean(), server_default=sa.text("false"), nullable=False
         ),
         sa.Column("room_mode", sa.String(), server_default="normal", nullable=False),
         sa.Column(

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/74b2b0236931_add_transcript_source_kind.py

@@ -20,11 +20,14 @@ depends_on: Union[str, Sequence[str], None] = None
 
 def upgrade() -> None:
     # ### commands auto generated by Alembic - please adjust! ###
+    sourcekind_enum = sa.Enum("room", "live", "file", name="sourcekind")
+    sourcekind_enum.create(op.get_bind())
+
     op.add_column(
         "transcript",
         sa.Column(
             "source_kind",
-            sa.Enum("ROOM", "LIVE", "FILE", name="sourcekind"),
+            sourcekind_enum,
             nullable=True,
         ),
     )
@@ -43,6 +46,8 @@ def upgrade() -> None:
 def downgrade() -> None:
     # ### commands auto generated by Alembic - please adjust! ###
     op.drop_column("transcript", "source_kind")
+    sourcekind_enum = sa.Enum(name="sourcekind")
+    sourcekind_enum.drop(op.get_bind())
 
 
 # ### end Alembic commands ###

server/migrations/versions/764ce6db4388_add_zulip_message_id.py

@@ -5,26 +5,28 @@ Revises: 62dea3db63a5
 Create Date: 2024-09-06 14:02:06.649665
 
 """
+
 from typing import Sequence, Union
 
-from alembic import op
 import sqlalchemy as sa
-
+from alembic import op
 
 # revision identifiers, used by Alembic.
-revision: str = '764ce6db4388'
-down_revision: Union[str, None] = '62dea3db63a5'
+revision: str = "764ce6db4388"
+down_revision: Union[str, None] = "62dea3db63a5"
 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.add_column('transcript', sa.Column('zulip_message_id', sa.Integer(), nullable=True))
+    op.add_column(
+        "transcript", sa.Column("zulip_message_id", sa.Integer(), nullable=True)
+    )
     # ### end Alembic commands ###
 
 
 def downgrade() -> None:
     # ### commands auto generated by Alembic - please adjust! ###
-    op.drop_column('transcript', 'zulip_message_id')
+    op.drop_column("transcript", "zulip_message_id")
     # ### end Alembic commands ###

server/migrations/versions/8120ebc75366_populate_webvtt_from_topics.py

@@ -0,0 +1,106 @@
+"""populate_webvtt_from_topics
+
+Revision ID: 8120ebc75366
+Revises: 116b2f287eab
+Create Date: 2025-08-11 19:11:01.316947
+
+"""
+
+import json
+from typing import Sequence, Union
+
+from alembic import op
+from sqlalchemy import text
+
+# revision identifiers, used by Alembic.
+revision: str = "8120ebc75366"
+down_revision: Union[str, None] = "116b2f287eab"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def topics_to_webvtt(topics):
+    """Convert topics list to WebVTT format string."""
+    if not topics:
+        return None
+
+    lines = ["WEBVTT", ""]
+
+    for topic in topics:
+        start_time = format_timestamp(topic.get("start"))
+        end_time = format_timestamp(topic.get("end"))
+        text = topic.get("text", "").strip()
+
+        if start_time and end_time and text:
+            lines.append(f"{start_time} --> {end_time}")
+            lines.append(text)
+            lines.append("")
+
+    return "\n".join(lines).strip()
+
+
+def format_timestamp(seconds):
+    """Format seconds to WebVTT timestamp format (HH:MM:SS.mmm)."""
+    if seconds is None:
+        return None
+
+    hours = int(seconds // 3600)
+    minutes = int((seconds % 3600) // 60)
+    secs = seconds % 60
+
+    return f"{hours:02d}:{minutes:02d}:{secs:06.3f}"
+
+
+def upgrade() -> None:
+    """Populate WebVTT field for all transcripts with topics."""
+
+    # Get connection
+    connection = op.get_bind()
+
+    # Query all transcripts with topics
+    result = connection.execute(
+        text("SELECT id, topics FROM transcript WHERE topics IS NOT NULL")
+    )
+
+    rows = result.fetchall()
+    print(f"Found {len(rows)} transcripts with topics")
+
+    updated_count = 0
+    error_count = 0
+
+    for row in rows:
+        transcript_id = row[0]
+        topics_data = row[1]
+
+        if not topics_data:
+            continue
+
+        try:
+            # Parse JSON if it's a string
+            if isinstance(topics_data, str):
+                topics_data = json.loads(topics_data)
+
+            # Convert topics to WebVTT format
+            webvtt_content = topics_to_webvtt(topics_data)
+
+            if webvtt_content:
+                # Update the webvtt field
+                connection.execute(
+                    text("UPDATE transcript SET webvtt = :webvtt WHERE id = :id"),
+                    {"webvtt": webvtt_content, "id": transcript_id},
+                )
+                updated_count += 1
+                print(f"✓ Updated transcript {transcript_id}")
+
+        except Exception as e:
+            error_count += 1
+            print(f"✗ Error updating transcript {transcript_id}: {e}")
+
+    print(f"\nMigration complete!")
+    print(f"  Updated: {updated_count}")
+    print(f"  Errors: {error_count}")
+
+
+def downgrade() -> None:
+    """Clear WebVTT field for all transcripts."""
+    op.execute(text("UPDATE transcript SET webvtt = NULL"))

server/migrations/versions/88d292678ba2_fix_transcript_json_nan_values.py

@@ -9,8 +9,6 @@ Create Date: 2025-07-15 19:30:19.876332
 from typing import Sequence, Union
 
 from alembic import op
-import sqlalchemy as sa
-
 
 # revision identifiers, used by Alembic.
 revision: str = "88d292678ba2"
@@ -21,7 +19,7 @@ depends_on: Union[str, Sequence[str], None] = None
 
 def upgrade() -> None:
     import json
-    import re
+
     from sqlalchemy import text
 
     # Get database connection
@@ -58,7 +56,9 @@ def upgrade() -> None:
             fixed_events = json.dumps(jevents)
             assert "NaN" not in fixed_events
         except (json.JSONDecodeError, AssertionError) as e:
-            print(f"Warning: Invalid JSON for transcript {transcript_id}, skipping: {e}")
+            print(
+                f"Warning: Invalid JSON for transcript {transcript_id}, skipping: {e}"
+            )
             continue
 
         # Update the record with fixed JSON

server/migrations/versions/9920ecfe2735_rename_transcript_to_text.py

@@ -5,13 +5,13 @@ Revises: 99365b0cd87b
 Create Date: 2023-11-02 18:55:17.019498
 
 """
+
 from typing import Sequence, Union
 
-from alembic import op
 import sqlalchemy as sa
-from sqlalchemy.sql import table, column
+from alembic import op
 from sqlalchemy import select
-
+from sqlalchemy.sql import column, table
 
 # revision identifiers, used by Alembic.
 revision: str = "9920ecfe2735"

server/migrations/versions/99365b0cd87b_add_title_short_and_long_summary_and_.py

@@ -8,8 +8,8 @@ Create Date: 2023-09-01 20:19:47.216334
 
 from typing import Sequence, Union
 
-from alembic import op
 import sqlalchemy as sa
+from alembic import op
 
 # revision identifiers, used by Alembic.
 revision: str = "99365b0cd87b"
@@ -22,7 +22,7 @@ def upgrade() -> None:
     # ### commands auto generated by Alembic - please adjust! ###
     op.execute(
         "UPDATE transcript SET events = "
-        'REPLACE(events, \'"event": "SUMMARY"\', \'"event": "LONG_SUMMARY"\');'
+        'REPLACE(events::text, \'"event": "SUMMARY"\', \'"event": "LONG_SUMMARY"\')::json;'
     )
     op.alter_column("transcript", "summary", new_column_name="long_summary")
     op.add_column("transcript", sa.Column("title", sa.String(), nullable=True))
@@ -34,7 +34,7 @@ def downgrade() -> None:
     # ### commands auto generated by Alembic - please adjust! ###
     op.execute(
         "UPDATE transcript SET events = "
-        'REPLACE(events, \'"event": "LONG_SUMMARY"\', \'"event": "SUMMARY"\');'
+        'REPLACE(events::text, \'"event": "LONG_SUMMARY"\', \'"event": "SUMMARY"\')::json;'
     )
     with op.batch_alter_table("transcript", schema=None) as batch_op:
         batch_op.alter_column("long_summary", nullable=True, new_column_name="summary")

server/migrations/versions/9f5c78d352d6_datetime_timezone.py

@@ -0,0 +1,121 @@
+"""datetime timezone
+
+Revision ID: 9f5c78d352d6
+Revises: 8120ebc75366
+Create Date: 2025-08-13 19:18:27.113593
+
+"""
+
+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 = "9f5c78d352d6"
+down_revision: Union[str, None] = "8120ebc75366"
+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(
+            "start_date",
+            existing_type=postgresql.TIMESTAMP(),
+            type_=sa.DateTime(timezone=True),
+            existing_nullable=True,
+        )
+        batch_op.alter_column(
+            "end_date",
+            existing_type=postgresql.TIMESTAMP(),
+            type_=sa.DateTime(timezone=True),
+            existing_nullable=True,
+        )
+
+    with op.batch_alter_table("meeting_consent", schema=None) as batch_op:
+        batch_op.alter_column(
+            "consent_timestamp",
+            existing_type=postgresql.TIMESTAMP(),
+            type_=sa.DateTime(timezone=True),
+            existing_nullable=False,
+        )
+
+    with op.batch_alter_table("recording", schema=None) as batch_op:
+        batch_op.alter_column(
+            "recorded_at",
+            existing_type=postgresql.TIMESTAMP(),
+            type_=sa.DateTime(timezone=True),
+            existing_nullable=False,
+        )
+
+    with op.batch_alter_table("room", schema=None) as batch_op:
+        batch_op.alter_column(
+            "created_at",
+            existing_type=postgresql.TIMESTAMP(),
+            type_=sa.DateTime(timezone=True),
+            existing_nullable=False,
+        )
+
+    with op.batch_alter_table("transcript", schema=None) as batch_op:
+        batch_op.alter_column(
+            "created_at",
+            existing_type=postgresql.TIMESTAMP(),
+            type_=sa.DateTime(timezone=True),
+            existing_nullable=True,
+        )
+
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("transcript", schema=None) as batch_op:
+        batch_op.alter_column(
+            "created_at",
+            existing_type=sa.DateTime(timezone=True),
+            type_=postgresql.TIMESTAMP(),
+            existing_nullable=True,
+        )
+
+    with op.batch_alter_table("room", schema=None) as batch_op:
+        batch_op.alter_column(
+            "created_at",
+            existing_type=sa.DateTime(timezone=True),
+            type_=postgresql.TIMESTAMP(),
+            existing_nullable=False,
+        )
+
+    with op.batch_alter_table("recording", schema=None) as batch_op:
+        batch_op.alter_column(
+            "recorded_at",
+            existing_type=sa.DateTime(timezone=True),
+            type_=postgresql.TIMESTAMP(),
+            existing_nullable=False,
+        )
+
+    with op.batch_alter_table("meeting_consent", schema=None) as batch_op:
+        batch_op.alter_column(
+            "consent_timestamp",
+            existing_type=sa.DateTime(timezone=True),
+            type_=postgresql.TIMESTAMP(),
+            existing_nullable=False,
+        )
+
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.alter_column(
+            "end_date",
+            existing_type=sa.DateTime(timezone=True),
+            type_=postgresql.TIMESTAMP(),
+            existing_nullable=True,
+        )
+        batch_op.alter_column(
+            "start_date",
+            existing_type=sa.DateTime(timezone=True),
+            type_=postgresql.TIMESTAMP(),
+            existing_nullable=True,
+        )
+
+    # ### end Alembic commands ###