MeetTrack: Intelligent Meeting Assistant

MeetTrack is a full-stack, AI-powered meeting assistant designed to securely transcribe audio, diarize speakers (who said what when), extract actionable tasks, and automatically email participants. It features a custom Python Playwright bot that can automatically join live Google Meet sessions and record them for you!

Key Features

1. Live Meeting Bot: A Python/Playwright headless bot that joins Google Meet, waits for admission, and records the meeting audio locally using ffmpeg and pulseaudio.
2. Automated Scheduling: Connect to your Google Calendar to fetch upcoming meetings, or manually schedule the bot. Celery Beat will automatically dispatch the bot exactly 2 minutes before the meeting starts.
3. AI Transcription & Diarization: High-speed speech-to-text using local faster-whisper and speaker identification using pyannote-audio, strictly constrained to run on a 4GB VRAM GPU without crashing.
4. Intelligent Task Extraction: Uses the Groq API (LLaMA-3) to read the transcript and extract action items, owners, and accurate deadlines.
5. Participant Mapping & Emails: Upload a CSV of participants, or pull them from Google Calendar. The system automatically sends notification emails with assigned tasks.
6. Queue System: A robust background worker queue (Celery + Redis) that restricts processing to one heavy AI task at a time (--concurrency=1) to prevent GPU Out-of-Memory (OOM) errors.

Technology Stack

• Frontend: React.js, Vite, Lucide Icons (Glassmorphic modern UI)
• Backend API: FastAPI (Python), SQLAlchemy, PostgreSQL
• Background Tasks: Celery, Celery Beat, Redis
• Live Bot: Python, Playwright, FFmpeg, PulseAudio
• AI Models: Faster-Whisper, PyAnnote, Groq API

---

System Requirements

• OS: Linux (Ubuntu/Debian recommended for PulseAudio and ffmpeg support)
• GPU: NVIDIA GPU (e.g., GTX 1050 Ti) with at least 4GB VRAM
• Dependencies: Node.js (v18+), Conda, FFmpeg, PulseAudio (sudo apt install ffmpeg pulseaudio)

---

Setup Instructions

1. Environment & Dependencies

Create the Conda environment and install all Python dependencies:

conda create -n meettrack python=3.11 -y
conda activate meettrack

# Install dependencies using the requirements file
pip install -r requirements.txt --extra-index-url https://download.pytorch.org/whl/cu121

# Install Playwright browsers (Required for the bot)
playwright install chromium

Install Frontend dependencies:

cd frontend
npm install
cd ..

2. Environment Variables (.env)

Create a .env file in the root directory:

# Database
DATABASE_URL=postgresql://<username>@localhost/meettrack

# AI Tokens
HUGGINGFACE_TOKEN=hf_your_token_here
GROQ_API_KEY=gsk_your_token_here

# Email Configuration (for sending task notifications)
SMTP_SERVER=smtp.gmail.com
SMTP_PORT=587
EMAIL_SENDER=[email protected]
EMAIL_APP_PASSWORD=your_app_password

# Google Calendar OAuth (Optional)
GOOGLE_CLIENT_ID=your_client_id
GOOGLE_CLIENT_SECRET=your_client_secret

Note: PyAnnote relies on a gated model. You MUST go to pyannote/speaker-diarization-community-1 and agree to their user conditions before the HuggingFace token will work.

3. Start Database and Redis

We run PostgreSQL and Redis directly from Conda for isolation:

conda install -c conda-forge postgresql redis -y
redis-server --daemonize yes
pg_ctl -D $CONDA_PREFIX/var/postgres -l $CONDA_PREFIX/var/postgres/server.log start

4. Database Migrations

Initialize the Postgres tables:

alembic upgrade head

---

Running the Application

You will need three terminal windows to run the full stack:

Terminal 1 (React Frontend):

cd frontend
npm run dev

Terminal 2 (FastAPI Server):

conda activate meettrack
uvicorn app.main:app --reload

Terminal 3 (Celery Worker + Scheduler):

conda activate meettrack
# The -B flag is CRITICAL! It enables the Celery Beat scheduler for automated bot dispatches.
celery -A app.worker worker -B --loglevel=info --concurrency=1

Critical: --concurrency=1 is required to ensure only 1 audio file is processed at a time, protecting the 4GB VRAM limit.

---

Usage

Open your browser to http://localhost:5173.

• Upload File: Drag and drop an existing .mp3 or .wav file to process it immediately.
• Send Live Bot: Paste a Google Meet link and schedule a time. The bot will automatically launch in the background 2 minutes before the meeting, wait for you to click "Ask to Join", and then silently record the audio!

---

Running with Docker (Alternative)

If you prefer to run the backend and worker in Docker, you can use docker-compose. Ensure you have Docker and NVIDIA Container Toolkit installed (you can use scripts/install_docker.sh to install these dependencies on Ubuntu).

# Build and start the services (Postgres, Redis, FastAPI, Celery Worker)
docker-compose up -d --build

# The frontend still needs to be run separately
cd frontend
npm run dev

Note: The worker container requires an NVIDIA GPU for AI models. Ensure nvidia-container-toolkit is installed on your host system.