Ledger — Meeting Intelligence Platform

Turn every meeting into a structured accountability record. Paste a transcript, upload a recording, or run your meeting inside Ledger — and get decisions, action items, risks, and a productivity score in under 30 seconds.

What It Does

Ledger solves a specific problem: meetings generate agreements and commitments that disappear the moment the call ends. Ledger captures those commitments and holds people accountable to them.

Capability	How it works
AI extraction	LLM reads transcript, outputs structured decisions, action items, and risks with owner names and confidence scores
Live meeting room	WebRTC peer-to-peer video/audio with in-browser transcription — transcript saved and auto-extracted on end
RAG (Ask AI)	Semantic search across all your meetings using vector embeddings; answer questions in plain English
Alerts	Automatic detection of unowned items, overdue tasks, unacknowledged assignments, and recurring risks
Metrics	Per-meeting productivity score and 4-week trend dashboard
Integrations	Import Zoom cloud recordings (audio or transcript); OAuth login via Google, Zoom, Microsoft, Slack
Notifications	HTML email for assignments, reminders, and weekly digests
Export	Branded PDF export of any meeting
Billing	Stripe subscription (Free / Pro)

Screenshots

Home	Dashboard

Meetings	Ask AI

Calendar	Integrations

System Architecture

┌──────────────────────────────────────────────────────────────────────┐
│                         CLIENT (Browser)                             │
│                                                                      │
│   React 19 + TypeScript + Tailwind CSS + React Router v7             │
│   Axios HTTP · WebSocket (chat, signaling) · WebRTC (video/audio)    │
│   Web Speech API (live captions) · MediaRecorder (audio capture)     │
└────────────────────────────┬─────────────────────────────────────────┘
                             │ HTTP / WS
                             ▼
┌──────────────────────────────────────────────────────────────────────┐
│                  NGINX (Reverse Proxy + Static Server)               │
│                                                                      │
│   / → serve React SPA (index.html fallback)                          │
│   /api/* → proxy to backend:8000                                     │
│   /ws/*  → WebSocket proxy (upgrade headers)                         │
│   Static assets cached 1 year (immutable)                            │
└────────────────────────────┬─────────────────────────────────────────┘
                             │
                             ▼
┌──────────────────────────────────────────────────────────────────────┐
│              BACKEND  (FastAPI + Uvicorn, 2 Gunicorn workers)        │
│                                                                      │
│  ┌─────────────────────────────────────────────────────────────┐    │
│  │  API Layer  (23 routers)                                    │    │
│  │  auth · meetings · transcripts · extract · decisions        │    │
│  │  action-items · risks · alerts · notifications · metrics    │    │
│  │  rag · live · rooms (WebRTC) · chat (WS) · upload           │    │
│  │  integrations · oauth · profile · colleagues · billing      │    │
│  │  export · messages                                          │    │
│  └───────────────────┬─────────────────────────────────────────┘    │
│                      │                                               │
│  ┌───────────────────▼─────────────────────────────────────────┐    │
│  │  Service Layer                                              │    │
│  │  openai_client   → selects OpenAI or Ollama                 │    │
│  │  ai_extractor    → prompt + JSON schema + confidence        │    │
│  │  rag             → chunk · embed · store · retrieve         │    │
│  │  transcription   → Whisper API or local whisper             │    │
│  │  email_notifier  → HTML templates, SMTP dispatch            │    │
│  │  zoom            → OAuth, recording list/download           │    │
│  └───────────────────┬─────────────────────────────────────────┘    │
│                      │                                               │
│  ┌───────────────────▼─────────────────────────────────────────┐    │
│  │  Worker Layer  (called synchronously, designed for async)   │    │
│  │  extract_from_transcript → idempotent extraction pipeline   │    │
│  │  alert_engine            → post-extraction alert detection  │    │
│  └─────────────────────────────────────────────────────────────┘    │
└──────────┬────────────────────────────────────────────────┬─────────┘
           │                                                │
           ▼                                                ▼
┌─────────────────────────┐               ┌────────────────────────────┐
│      MySQL 8.0          │               │  ChromaDB  (vector store)  │
│                         │               │                            │
│  users                  │               │  Collection: meeting_chunks│
│  meetings               │               │  Embeddings: all-MiniLM-   │
│  transcripts            │               │             L6-v2 (384-dim)│
│  decisions              │               │  Distance:  cosine         │
│  action_items           │               │  Metadata:  meeting_id,    │
│  risks                  │               │             title, date    │
│  alerts                 │               │  Chunks:    500 words,     │
│  meeting_participants   │               │             50-word overlap│
│  colleagues             │               └────────────────────────────┘
│  messages               │
└─────────────────────────┘

                External Services
                ─────────────────
                OpenAI API     → gpt-4o-mini (extraction, live assist)
                               → whisper-1 (audio transcription)
                Ollama         → local LLM fallback (llama3, mistral, etc.)
                Stripe         → checkout, portal, webhooks
                Zoom API       → OAuth, recording list/download/transcript
                Google OAuth   → sign in with Google
                Microsoft OAuth→ sign in with Microsoft
                Slack OAuth    → sign in with Slack
                SMTP           → email notifications

Data Model

users
├── id            UUID PK
├── name          VARCHAR(255)
├── email         VARCHAR(255) UNIQUE
├── role          VARCHAR(255)
├── password_hash VARCHAR(255)          # null for OAuth-only accounts
├── avatar_url    VARCHAR(500)
├── stripe_customer_id    VARCHAR(255)
├── stripe_subscription_id VARCHAR(255)
├── password_reset_token  VARCHAR(255)
├── password_reset_expires DATETIME
└── created_at    DATETIME

meetings
├── id            UUID PK
├── title         VARCHAR(500)
├── platform      VARCHAR(100)         # Zoom | Teams | Meet | Ledger | etc.
├── start_time    DATETIME
├── end_time      DATETIME
├── owner_id      UUID → users.id
└── created_at    DATETIME

transcripts
├── id            UUID PK
├── meeting_id    UUID → meetings.id   # effectively 1:1
├── content       TEXT                 # raw transcript
└── created_at    DATETIME

decisions
├── id            UUID PK
├── meeting_id    UUID → meetings.id
├── owner_id      UUID → users.id      # nullable
├── summary       TEXT
├── source_sentence TEXT               # excerpt from transcript
├── confidence    FLOAT                # 0.0–1.0, LLM self-reported
└── created_at    DATETIME

action_items
├── id            UUID PK
├── meeting_id    UUID → meetings.id
├── owner_id      UUID → users.id      # nullable
├── description   TEXT
├── status        VARCHAR(50)          # open | done
├── due_date      DATETIME
├── source_sentence TEXT
├── confidence    FLOAT
├── acknowledged_at DATETIME           # when owner first saw it
└── created_at    DATETIME

risks
├── id            UUID PK
├── meeting_id    UUID → meetings.id
├── description   TEXT
├── source_sentence TEXT
├── confidence    FLOAT
└── created_at    DATETIME

alerts
├── id            UUID PK
├── meeting_id    UUID → meetings.id
├── action_item_id UUID → action_items.id
├── type          VARCHAR(100)         # see Alert Types below
├── message       TEXT
└── created_at    DATETIME

meeting_participants
├── meeting_id    UUID → meetings.id   # composite PK
├── user_id       UUID → users.id      # composite PK
└── role          VARCHAR(100)         # presenter | attendee | organizer

colleagues
├── id            UUID PK
├── user_id       UUID → users.id ON DELETE CASCADE
├── colleague_id  UUID → users.id ON DELETE CASCADE
└── created_at    DATETIME             # UNIQUE(user_id, colleague_id)

messages
├── id            UUID PK
├── sender_id     UUID → users.id ON DELETE CASCADE
├── recipient_id  UUID → users.id ON DELETE CASCADE
├── content       TEXT
├── read_at       DATETIME
└── created_at    DATETIME

Alert Types

Type	Trigger
`no_owner`	Action item extracted with no owner
`decision_no_owner`	Decision extracted with no owner
`overdue`	Action item past due date and not done
`no_outcomes`	Meeting had no decisions and no actions
`never_acknowledged`	Action assigned 2+ days ago, never acknowledged
`repeated_issue`	Risk appeared in 2+ meetings in the last 30 days (≥3-word overlap)

Backend Services & Workers

AI Extractor (`services/ai_extractor.py`)

Sends the transcript to an LLM with a structured prompt that instructs it to return a strict JSON schema:

{
  "decisions":    [{ "summary", "owner", "source_sentence", "confidence" }],
  "action_items": [{ "description", "owner", "due_date", "source_sentence", "confidence" }],
  "risks":        [{ "description", "source_sentence", "confidence" }]
}

Priority: OpenAI gpt-4o-mini → Ollama (local fallback). The USE_OLLAMA=true env var skips OpenAI entirely.

RAG Engine (`services/rag.py`)

Implements semantic search across all meeting transcripts.

Indexing:

Chunk transcript into 500-word windows with 50-word overlap
Embed each chunk with sentence-transformers/all-MiniLM-L6-v2 (384-dim)
Upsert into ChromaDB collection meeting_chunks with metadata

Querying:

Embed the user's question
Cosine-similarity search in ChromaDB (top_k=5)
Build context string from retrieved chunks
Pass context + question to Ollama → return grounded answer

Auto-triggered after every extraction. Can also be triggered manually via /rag/index-all.

Extraction Worker (`workers/extract_from_transcript.py`)

Idempotent pipeline — safe to re-run on the same transcript:

1. DELETE existing decisions, action_items, risks for meeting_id
2. Call extract_decisions_and_actions(llm, transcript.content)
3. For each result: match owner_name to existing User.name (no ghost accounts)
4. Save decisions / action items / risks with confidence + source_sentence
5. run_alerts_for_meeting(db, meeting_id)
6. detect_repeated_issues(db, meeting_id)

Alert Engine (`workers/alert_engine.py`)

Runs after every extraction. Clears old alerts first (idempotent), then checks:

Action items without an owner
Decisions without an owner
Action items that are overdue
Meetings with zero outcomes
Action items unacknowledged for 48+ hours
Risks matching ≥3 words from any risk in the past 30 days of the user's meetings

Transcription (`services/transcription.py`)

Accepts .mp3 .mp4 .mpeg .mpga .m4a .wav .webm .ogg up to 25 MB. Tries OpenAI Whisper API first; falls back to local whisper if use_local_whisper=true.

Email Notifier (`services/email_notifier.py`)

Four HTML-templated email types, all using SMTP with STARTTLS:

Function	Trigger
`send_action_item_assigned`	Owner assigned via extraction
`send_action_item_reminder`	Overdue / due today / due in N days
`send_weekly_digest`	Stats + open items list
`send_meeting_summary`	Post-extraction summary to meeting owner

All links resolve to FRONTEND_URL from env.

API Reference

Authentication

Method	Path	Auth	Description
POST	`/auth/signup`	—	Register with email + password
POST	`/auth/login`	—	Returns JWT
GET	`/oauth/{provider}/login`	—	Redirect to OAuth consent screen
GET	`/oauth/{provider}/callback`	—	Exchange code → JWT, create user if new
POST	`/auth/forgot-password`	—	Send reset token
POST	`/auth/reset-password`	—	Consume token, set new password

Meetings

Method	Path	Auth	Description
GET	`/meetings/`	✓	List — query params: `q` (search), `platform`
POST	`/meetings/`	✓	Create
GET	`/meetings/{id}`	✓	Full detail (transcript, decisions, actions, risks)
PUT	`/meetings/{id}`	✓	Update title/times
GET	`/meetings/calendar`	✓	`year`, `month` query params
POST	`/meetings/{id}/participants`	✓	Add participant
DELETE	`/meetings/{id}/participants/{uid}`	✓	Remove participant

Transcript & Upload

Method	Path	Auth	Description
POST	`/transcripts/`	✓	Paste raw transcript text
POST	`/upload/audio`	✓	Upload audio file → auto-transcribed (max 25 MB)

Extraction & Outcomes

Method	Path	Auth	Description
POST	`/extract/`	—	Run AI extraction on a transcript
GET	`/alerts/{meeting_id}`	✓	Get alerts for a meeting
POST	`/action-items/{id}/acknowledge`	✓	Mark as seen
POST	`/action-items/{id}/done`	✓	Mark complete
POST	`/action-items/{id}/reopen`	✓	Reopen

RAG (Ask AI)

Method	Path	Auth	Description
POST	`/rag/query`	—	Semantic Q&A across all meetings
GET	`/rag/stats`	✓	Index stats (chunks, meetings, health)
POST	`/rag/index-all`	—	Re-index all transcripts
DELETE	`/rag/clear`	—	Clear index

Live Meeting Room

Method	Path	Auth	Description
POST	`/rooms/create`	✓	Create room → returns 8-char `room_id`
GET	`/rooms/{room_id}`	✓	Room info + participant list
POST	`/rooms/{room_id}/end`	✓	Save transcript + trigger extraction
WS	`/ws/room/{room_id}?token=`	✓	WebRTC signaling (offer/answer/ICE/chat/reaction)
WS	`/ws/chat?token=`	✓	Real-time 1:1 messaging

Metrics & Export

Method	Path	Auth	Description
GET	`/metrics/dashboard`	✓	KPIs, weekly trend, recent alerts
GET	`/metrics/meeting/{id}`	✓	Per-meeting productivity score
GET	`/export/meeting/{id}/pdf`	✓	Download branded PDF

Integrations & Billing

Method	Path	Auth	Description
GET	`/integrations/zoom/status`	—	Is Zoom configured?
GET	`/integrations/zoom/recordings`	✓	List cloud recordings
POST	`/integrations/zoom/import`	✓	Import a recording
GET	`/billing/status`	✓	Free or Pro
POST	`/billing/create-checkout-session`	✓	Start Stripe checkout
POST	`/billing/portal`	✓	Stripe customer portal
POST	`/billing/webhook`	—	Stripe event handler

Data Flow

Transcript → Outcomes

User pastes transcript / uploads audio
          │
          ▼
POST /transcripts  or  POST /upload/audio → Whisper transcription
          │
          ▼
POST /extract
          │
          ├─► Delete old decisions/actions/risks (idempotent)
          │
          ├─► LLM (gpt-4o-mini or Ollama)
          │       Returns JSON: decisions[], action_items[], risks[]
          │       Each item has: owner, source_sentence, confidence
          │
          ├─► Owner matching: User.name lookup (no ghost accounts)
          │
          ├─► Save to DB
          │
          ├─► Alert engine
          │       no_owner, overdue, no_outcomes, unacknowledged,
          │       decision_no_owner, repeated_issue
          │
          ├─► RAG indexing (async, best-effort)
          │       Chunk → embed → upsert ChromaDB
          │
          └─► Email notifications (async, best-effort)

Semantic Q&A (RAG)

User: "What did we decide about the API redesign?"
          │
          ▼
POST /rag/query
          │
          ├─► Embed question (all-MiniLM-L6-v2, 384-dim)
          │
          ├─► ChromaDB cosine search → top 5 chunks
          │       Each chunk: text excerpt + {meeting_id, title, date}
          │
          ├─► Build context string
          │
          └─► Ollama generates grounded answer
                  Returns: { answer, sources[] }

Live Meeting Room (WebRTC)

User A clicks "Start Meeting" → names meeting
          │
          ▼
POST /rooms/create → room_id (8-char)
          │
          ▼
Browser opens /room/{room_id}
          │
          ├─► WS connect: /ws/room/{room_id}?token=
          │
          ├─► Peer joins → receive peer_joined (should_offer: true/false)
          │       → RTCPeerConnection offer/answer via WS signaling
          │       → ICE candidates exchanged
          │       → P2P audio/video established (STUN: Google)
          │
          ├─► Web Speech API running in browser
          │       Real-time captions displayed on screen
          │       Transcript accumulated in memory
          │
          └─► User clicks "End Call"
                  │
                  ▼
                POST /rooms/{room_id}/end
                  { title, transcript, participants[], start_time, end_time }
                  │
                  ▼
                Create Meeting + Transcript records
                  │
                  ▼
                Run extraction pipeline (async)
                  │
                  ▼
                Navigate to /meetings/{id}

Alert Generation

After extraction:
  run_alerts_for_meeting(meeting_id)
    │
    ├─► Check each action item: no owner? → alert
    ├─► Check each action item: past due? → alert
    ├─► Check each decision: no owner? → alert
    ├─► No decisions AND no actions? → alert
    ├─► Action assigned 48h+ ago, not acknowledged? → alert
    │
    └─► detect_repeated_issues(meeting_id)
          │
          ├─► Get current risks
          ├─► Query past 30 days of user's meetings
          └─► Word overlap ≥3 with any past risk → repeated_issue alert

Frontend Architecture

Built with React 19, TypeScript, React Router v7, and Tailwind CSS. The entire app is a single-page application served by nginx; all routing is client-side.

Route Map

/                    → Home (landing, pricing)
/login               → Email/OAuth login
/signup              → Registration
/forgot-password     → Password reset request
/reset-password      → Password reset confirm
/oauth-callback      → Extract JWT from URL, redirect
/meetings            → Meeting list + search + filters
/meetings/:id        → Full detail: transcript, outcomes, participants, export
/dashboard           → KPI cards, 4-week trend chart, recent alerts
/calendar            → Monthly calendar view grouped by meeting
/chat                → Real-time 1:1 messaging
/integrations        → Zoom/Teams/Meet config + Zoom import UI
/profile             → Avatar, name, password, billing
/demo                → Pre-populated example meeting
/room/:roomId        → Live WebRTC meeting room

State Management

No external state library — local useState / useEffect per page, JWT in localStorage, Axios instance with auth header injected from lib/api.ts.

Real-time

Feature	Protocol	Notes
1:1 chat	WebSocket `/ws/chat`	ConnectionManager in-memory on server
Meeting room signaling	WebSocket `/ws/room/{id}`	Offer/answer/ICE, chat, reactions, media state
WebRTC audio/video	P2P (browser↔browser)	Google STUN servers; no media relay (TURN)
Live captions	Web Speech API	Runs entirely in browser, no server round-trip

Theming

Dark mode (default) and light mode toggled by .light class on <html>. All Tailwind slate-* values have explicit light-mode overrides in index.css. Accent color: ledger-pink (#e485b6).

Deployment

Docker Compose (recommended)

git clone https://github.com/pragyabose1011/Ledger.git
cd Ledger

cp backend/.env.example backend/.env
# Edit backend/.env — set SECRET_KEY and OPENAI_API_KEY at minimum

docker compose up --build

App: http://localhost
API docs: http://localhost:8000/docs

Services:

Container	Image	Port	Volumes
`ledger-db`	mysql:8.0	internal	`mysql-data`
`ledger-backend`	custom (python:3.11-slim)	8000	`chroma-data`, `uploads-data`, `model-cache`
`ledger-frontend`	custom (nginx:alpine)	80	—

mysql-data, chroma-data, and uploads-data are Docker named volumes — data persists across restarts. Only rebuild the backend after backend changes:

docker compose up --build backend -d

Without Docker

Backend

cd backend
python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt
uvicorn app.main:app --reload

Frontend

cd frontend
npm install
npm run dev

MySQL must be running separately. Set DATABASE_URL in backend/.env.

Environment Variables

Create backend/.env:

# ── Required ────────────────────────────────────────────
JWT_SECRET_KEY=<long-random-string>
DATABASE_URL=mysql+pymysql://ledger:ledgerpass@localhost:3306/ledger

# ── AI — pick one ────────────────────────────────────────
OPENAI_API_KEY=sk-...
# OR
USE_OLLAMA=true
OLLAMA_MODEL=llama3.2
OLLAMA_BASE_URL=http://localhost:11434

# ── Frontend URL (controls OAuth redirects + email links) ─
FRONTEND_URL=http://localhost

# ── OAuth providers (all optional) ───────────────────────
GOOGLE_CLIENT_ID=
GOOGLE_CLIENT_SECRET=
ZOOM_CLIENT_ID=
ZOOM_CLIENT_SECRET=
ZOOM_ACCOUNT_ID=
MICROSOFT_CLIENT_ID=
MICROSOFT_CLIENT_SECRET=
SLACK_CLIENT_ID=
SLACK_CLIENT_SECRET=

# ── Email (optional) ─────────────────────────────────────
SMTP_HOST=smtp.gmail.com
SMTP_PORT=587
SMTP_USER=you@gmail.com
SMTP_PASS=your-app-password
EMAIL_FROM_NAME=Ledger

# ── Stripe (optional) ────────────────────────────────────
STRIPE_SECRET_KEY=sk_live_...
STRIPE_PRICE_ID=price_...
STRIPE_WEBHOOK_SECRET=whsec_...

# ── Advanced ─────────────────────────────────────────────
EXTRA_CORS_ORIGINS=https://your-domain.com   # comma-separated additional origins

Tech Stack

Layer	Technology	Version
Frontend framework	React	19
Language	TypeScript	5.9
Styling	Tailwind CSS	3.4
Routing	React Router	7
HTTP client	Axios	1.13
Build tool	Vite	7
Backend framework	FastAPI	0.104
ASGI server	Uvicorn / Gunicorn	—
ORM	SQLAlchemy	2.0
Database	MySQL	8.0
Auth	JWT (python-jose) + OAuth2	HS256 / 7-day
Password hashing	passlib (pbkdf2_sha256)	—
LLM	OpenAI gpt-4o-mini / Ollama	—
Transcription	OpenAI Whisper / local whisper	—
Vector store	ChromaDB	0.4
Embeddings	sentence-transformers all-MiniLM-L6-v2	384-dim
PDF generation	ReportLab	4
Payments	Stripe	7
WebRTC signaling	FastAPI WebSocket	—
Live captions	Web Speech API	browser native
Proxy	nginx	alpine
Container orchestration	Docker Compose	3.8

Security

Area	Implementation
Password policy	8+ chars, uppercase, lowercase, digit, special char enforced server-side
Disposable email blocking	20+ known throwaway domains rejected at signup
JWT	HS256, 7-day expiry, warning logged if default dev secret is used
CORS	Configurable via `FRONTEND_URL` + `EXTRA_CORS_ORIGINS`; defaults to localhost only
SQL injection	Not possible — all queries via SQLAlchemy ORM
File upload	Extension allowlist, 25 MB size cap, saved with UUID filenames
OAuth	Minimal scopes; tokens not stored in DB (used only to create/find user)
Stripe webhooks	Signature verification (`stripe.Webhook.construct_event`) before processing
Data isolation	All metrics, meetings, and action items filtered by `owner_id = current_user.id`
Ownership enforcement	Action item mutations verify parent meeting belongs to current user

License

MIT

Name		Name	Last commit message	Last commit date
.github		.github
.vscode		.vscode
backend		backend
docs/screenshots		docs/screenshots
extension		extension
frontend		frontend
infra/k8s		infra/k8s
.env.example		.env.example
.gitignore		.gitignore
LICENSE		LICENSE
Makefile		Makefile
README.md		README.md
docker-compose.yml		docker-compose.yml

Latest commit

History

Folders and files

Repository files navigation