Dozi

AI-powered sales conversation assistant that provides real-time insights during calls, meetings, and interviews.

Architecture

Each service is independently deployed — there is no monolithic setup.

User's Browser
  ├── Supabase (direct)          — auth, conversations, transcriptions, battle_cards, preferences
  ├── LLM Service (FastAPI)      — model registry, LiveKit tokens, battlecard generation, prompts
  └── LiveKit Cloud (WebSocket)  — audio streaming
       └── Transcription Agent  — STT via Deepgram/OpenAI
            └── LLM Service     — GET /models at startup

The LLM Service is stateless and has no database connection — it handles battle card generation, LiveKit room/token management, and serves the model registry. The Transcription Agent runs in LiveKit Cloud and handles all STT. Supabase is accessed directly from the frontend for all persistence.

Tech Stack

• LLM Service: FastAPI, Python 3.13+, uv
• Frontend: React 18, TypeScript, Vite, pnpm
• Database: Supabase (PostgreSQL + Auth)
• Audio/Transcription: LiveKit Cloud, Deepgram Nova-3 (streaming STT)
• AI: Groq (battle cards), OpenAI (fallback)

Setup

1. Supabase

Requires Node.js (for npx). Create a project at https://supabase.com, then run the migration:

npx supabase login
npx supabase link
npx supabase db push

2. LLM Service

cd llm_service
cp .env.example .env
# Fill in: OPENAI_API_KEY, LIVEKIT_*, AUTH_JWKS_URL values

Option A: Direct (recommended for development)

uv sync
PYTHONPATH=. uv run uvicorn llm_service.main:app --reload

Option B: Docker

docker compose up -d --build   # from llm_service/

LLM service runs at http://localhost:8000

3. Transcription Agent

Install LiveKit CLI:

brew install livekit-cli                          # macOS
curl -sSL https://get.livekit.io/cli | bash       # Linux

Deploy agent:

lk cloud auth
cd agents/
lk agent create --project dozi

Set DEEPGRAM_API_KEY (and optionally OPENAI_API_KEY) in LiveKit Cloud dashboard under Settings → Agents.

For local development, see agents/README.md.

4. Frontend

cd frontend
pnpm install

cp apps/web/.env.example apps/web/.env
# Fill in: VITE_SUPABASE_URL, VITE_SUPABASE_PUBLISHABLE_KEY from step 1

pnpm dev

Frontend runs at http://localhost:5173

Project Structure

dozi/
├── agents/                     # LiveKit transcription agent (Deepgram/OpenAI)
│   ├── transcription_agent.py
│   ├── requirements.txt
│   └── Dockerfile
├── llm_service/                # FastAPI LLM service (self-contained)
│   ├── llm_service/            # Python package
│   │   ├── api/routes/         # battlecards, livekit, models
│   │   ├── prompts/            # Prompt loader + models
│   │   ├── services/           # Business logic
│   │   ├── schemas.py          # All Pydantic schemas
│   │   ├── auth.py             # JWT validation (Supabase JWKS)
│   │   └── settings.py         # Env config + model registry
│   ├── prompts/                # Prompt YAML files (call, meeting, interview)
│   ├── models.yaml             # Single source of truth for all model names
│   ├── Dockerfile
│   └── docker-compose.yml
├── frontend/                   # React frontend (pnpm monorepo)
│   ├── apps/web/               # Main web app
│   └── packages/api-client/   # @dozi/api-client
└── supabase/                   # Database migrations

Deployment

Each service deploys independently:

• Frontend — deploy frontend/apps/web/ to Vercel, Netlify, or any static host
• LLM Service — deploy as a Docker container on any VPS

# LLM Service on VPS
git clone <repo> /opt/dozi && cd /opt/dozi/llm_service
cp .env.example .env
# Fill in real values, set CORS_ORIGINS=https://your-frontend-domain.com

docker compose up -d --build

Update with git pull && docker compose up -d --build (from llm_service/).

• Transcription Agent — deploy to LiveKit Cloud via lk agent create
• Supabase — managed service, run supabase db push for migrations

Development

This project uses pre-commit hooks for code quality (Ruff + ty + ESLint + tsc). Install from the repo root:

uv sync --dev    # from llm_service/ for the uv environment
pre-commit install

License

AGPL-3.0