Mira is a sophisticated AI-powered personal assistant that combines real-time voice interaction, custom-trained language models, and a modern desktop interface. Built with a focus on privacy, performance, and extensibility, Mira represents a complete end-to-end solution for voice-based AI interaction.

• Real-time Voice Interaction: Powered by Whisper Live technology for instant speech-to-text conversion
• Custom-trained AI Models: Two specialized language models fine-tuned for specific tasks
• Cross-platform Desktop Client: Modern Electron-based interface for macOS and Windows
• Secure Network Architecture: Private deployment using Tailscale for device connectivity
• Extensible Backend: FastAPI-based Python backend with modular architecture
• Comprehensive Testing: 67+ tests across all components

• Fine-tuned version of LLaMA-2-7B-Chat optimized for:
  • Natural language command interpretation
  • Wake word response handling
  • Function calling and system control
• Utilizes LoRA (Low-Rank Adaptation) for efficient training
• Optimized for real-time response generation

• Custom-tuned version of TII UAE Falcon-40B-Instruct for:
  • Structured data extraction from speech
  • Contact information parsing
  • Calendar event recognition
  • Task and reminder extraction
• Enhanced with task-specific prompts and examples
• Optimized for accuracy in entity extraction

• Core Components:

  • FastAPI application server
  • SQLite database with SQLAlchemy ORM
  • Custom ML model manager
  • Real-time audio processing pipeline

• API Routes:

  /interactions
  ├── POST /register          # Register new voice interactions
  ├── GET /{id}              # Retrieve specific interactions
  ├── POST /{id}/inference   # Run inference on interaction
  └── DELETE /{id}           # Remove interaction
  
  /conversations
  └── GET /all               # Retrieve conversation history
  
  /persons
  └── [Person management endpoints]
  
  /services
  └── [System service endpoints]
  
  /streams
  └── [Real-time stream management]
  

• Modern UI with intuitive design
• Real-time voice capture and streaming
• Secure IPC bridge architecture
• Professional animations and visual feedback
• Keyboard shortcuts for efficiency

# Clone the repository
git clone https://github.com/your-org/mira-assistant.git
cd mira-assistant/backend

# Create and activate virtual environment
python -m venv .venv
source .venv/bin/activate  # On Windows: .venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

# Start the server
python mira.py

# Navigate to desktop client
cd ../desktop-client

# Install dependencies
npm install

# Start the application
npm start

# For development mode
npm run start-dev

Mira uses Tailscale for secure, private networking between devices:

1. Network Setup:

   • Install Tailscale on all devices
   • Create a private Tailscale network
   • Configure subnet routing for local services

2. Backend Configuration:

   • Deploy backend on primary device
   • Expose backend port through Tailscale network
   • Configure client to use Tailscale IP

3. Client Configuration:

   • Update baseUrl in client config to use Tailscale IP
   • Enable secure WebSocket connections
   • Maintain end-to-end encryption

# Backend Tests
cd backend
python -m pytest tests/

# Desktop Client Tests
cd desktop-client
npm test

cd backend/tuning

# Generate training datasets
python acquire_datasets.py --task both --output-dir datasets/

# Fine-tune models
python fine_tune_models.py --model llama-2-7b-chat-hf-function-calling-v3
python fine_tune_models.py --model tiiuae-falcon-40b-instruct

cd desktop-client

# Build for macOS
npm run build-mac

# Build for Windows
npm run build-win

# Build for all platforms
npm run build

• Backend:

  • Python 3.8+
  • FastAPI
  • SQLAlchemy
  • PyTorch
  • Transformers
  • Whisper Live

• Frontend:

  • Electron
  • Node.js
  • Modern JavaScript (ES6+)
  • HTML5/CSS3
  • Web Audio API

• ML/AI:

  • LLaMA-2-7B-Chat
  • Falcon-40B-Instruct
  • LoRA fine-tuning
  • Custom training pipelines

• Deployment:

  • Tailscale
  • SQLite
  • LM Studio
  • Electron Builder

Contributions are welcome! Please read our Contributing Guidelines for details on our code of conduct and the process for submitting pull requests.

This project is licensed under the MIT License - see the LICENSE file for details.

• The LLaMA team at Meta AI Research
• The Falcon team at Technology Innovation Institute
• The Whisper team at OpenAI
• The Tailscale team for their excellent networking solution