🎙️ AI Meeting Tracker

Self-hosted AI-powered meeting transcription and task extraction

✨ Features

• 🎯 Upload recordings (MP3, MP4, WAV, M4A) - up to 1.5 hours
• 🎙️ Transcribe with Whisper (local, no API costs)
• 🤖 Extract tasks with Ollama (Llama 3.1 or Mistral)
• 💬 Chat about meetings - ask questions about transcript
• 📚 Meeting history - last 10 meetings in sidebar
• 💾 Database persistence - SQLite storage for all meetings
• ✅ Structured tasks with owner, deadline, confidence scores
• 🔒 100% self-hosted - no cloud dependencies, privacy-first

🛠️ Tech Stack

• Backend: Python 3.11+ with FastAPI
• Database: SQLite with SQLAlchemy ORM
• Transcription: OpenAI Whisper (100% local, no API)
• AI Extraction: Ollama (Llama 3.1 8B)
• Frontend: Vue 3 + Vite + Vue Router
• Styling: Custom CSS with glassmorphism design

📋 System Requirements

• Python: 3.11 or higher
• Node.js: 18+
• RAM: 10GB+ (for Whisper + Ollama)
• Disk: ~5GB for AI models
• GPU: Optional (CPU works fine, just slower)
• OS: Linux, macOS, Windows (WSL)
• FFmpeg: Required for audio processing

🚀 Quick Start

0. Install FFmpeg (Required)

# Ubuntu/Debian
sudo apt-get install ffmpeg

# macOS
brew install ffmpeg

# Check installation
ffmpeg -version

1. Install Ollama

chmod +x scripts/install_ollama.sh
./scripts/install_ollama.sh

This will:

• Install Ollama
• Pull Llama 3.1 8B model (~4.7GB)

2. Setup Backend

chmod +x setup_backend.sh
./setup_backend.sh

3. Setup Frontend

chmod +x setup_frontend.sh
./setup_frontend.sh

4. Run the Application

Terminal 1 - Start Backend:

cd backend
source venv/bin/activate
python main.py

Backend runs on: http://localhost:8000

Terminal 2 - Start Frontend:

cd frontend
npm run dev

Frontend runs on: http://localhost:5173

5. Test It

1. Open http://localhost:5173 in your browser
2. Upload a meeting recording
3. Click "Process Recording"
4. Wait for transcription + AI extraction (~1-2 minutes)
5. View extracted tasks with confidence scores!

📖 How It Works

1. Upload → File saved to backend/uploads/
2. Transcribe → OpenAI Whisper converts full audio to text (up to 1.5 hours)
3. Extract → Ollama (Llama 3.1) identifies action items from transcript
4. Store → Meeting, tasks, and metadata saved to SQLite database
5. Display → Tasks shown with owner, deadline, and confidence scores
6. Chat → Ask questions about the meeting using AI

📁 Project Structure

ProjectX/
├── backend/
│   ├── main.py                    # FastAPI application
│   ├── database.py                # SQLAlchemy models
│   ├── config.py                  # Environment configuration
│   ├── requirements.txt           # Python dependencies
│   ├── modules/
│   │   ├── transcription.py      # Whisper integration
│   │   └── task_extractor.py     # Ollama task extraction
│   ├── uploads/                   # Temporary file storage
│   └── meetings.db                # SQLite database
├── frontend/
│   ├── src/
│   │   ├── App.vue               # Main layout with history sidebar
│   │   ├── pages/
│   │   │   ├── UploadPage.vue    # File upload interface
│   │   │   └── ResultsPage.vue   # Results with chat
│   │   ├── composables/
│   │   │   ├── useHistory.js     # LocalStorage history
│   │   │   └── useSession.js     # Session management  
│   │   ├── router/
│   │   │   └── index.js          # Vue Router config
│   │   ├── style.css             # Modern dark theme
│   │   └── main.js               # Vue entry point
│   └── package.json
├── scripts/
│   └── install_ollama.sh         # Ollama installation
├── .env.example                   # Environment config template
├── .gitignore
├── LICENSE                        # MIT License
└── README.md

🔧 Configuration

Change AI Model

Edit backend/modules/task_extractor.py:

async def extract_tasks(transcript: str, model: str = "mistral:7b"):
    # Change model here

Available models:

• llama3.1:8b (default, best accuracy)
• mistral:7b (faster, good accuracy)
• llama2:7b (lighter)

Pull new models: ollama pull mistral:7b

Change Whisper Model

Edit backend/modules/transcription.py:

_model = WhisperModel("medium", device="cpu", compute_type="int8")

Models: tiny, base, small, medium, large

🐛 Troubleshooting

Ollama connection error:

# Check if Ollama is running
ollama list

# Start Ollama service
ollama serve

Whisper model download:

• First run will download Whisper model (~150MB for base)
• This only happens once

Port already in use:

# Change port in backend/main.py
uvicorn.run(app, host="0.0.0.0", port=8001)

🎯 Roadmap

Current Status: Working Prototype with Database ✅

Completed:

• ✅ SQLite database persistence
• ✅ Meeting history and retrieval
• ✅ Chat feature for Q&A about meetings
• ✅ Full audio transcription (up to 1.5 hours)
• ✅ Vue Router with separate pages
• ✅ Modern UI with glassmorphism
• ✅ Health check API endpoint

Coming Soon (Production Ready):

• 🚧 Docker containerization
• 🚧 Production frontend build
• 🚧 Security hardening (rate limiting, file validation)
• 🚧 Deployment documentation

Future MVP Features:

• 📋 Task editing (CRUD operations)
• 👥 Speaker diarization (identify who said what)
• 📊 Kanban board integration
• 🎨 Enhanced UI animations
• 🧪 Automated tests

📝 License

MIT

Built with:

• faster-whisper - Optimized Whisper
• Ollama - Local LLM runtime
• FastAPI - Modern Python API
• Vue 3 - Progressive framework

---

Made with ❤️ for productivity nerds who hate cloud subscriptions