Nova - Privacy-First Browser Activity Monitor

Nova is an intelligent, privacy-preserving browser activity monitor that leverages hybrid AI to analyze your digital habits. By combining the speed and privacy of on-device inference (Cactus + FunctionGemma) with the reasoning capabilities of the cloud (Gemini), Nova provides real-time insights without compromising user data.

Features

• Privacy-First Analysis: Your browsing data is analyzed locally on your device by default. Sensitive information never leaves your machine unless complex reasoning is required.
• Hybrid Intelligence: Seamlessly switches between a lightweight local model (FunctionGemma 270M) for speed and a powerful cloud model (Gemini 3.0 Flash) for complex tasks.
• Real-Time Categorization: Instantly categorizes web activity (e.g., Coding, Social Media, Productivity) and assesses productivity risk.
• Chrome Extension: A lightweight browser extension captures events and provides a live feed of your digital footprint.
• Optimized for Apple Silicon: Built on the Cactus engine, delivering up to 3000 tokens/sec prefill speed on M-series chips.

Technologies Used

• On-Device AI: Google FunctionGemma (270M parameters) running on Cactus.
• Cloud AI: Google Gemini 3.0 Flash for fallback reasoning.
• Backend: Python, FastAPI, Uvicorn.
• Frontend: Chrome Extension (Manifest V3), HTML/JS/CSS.
• Tooling: cactus-python SDK, google-genai SDK.

Project Structure

├── monitor-backend/        # 🧠 The Brain
│   ├── main.py             # FastAPI server & AI orchestration
│   ├── tools/              # AI Tool definitions (API, Context)
│   └── storage/            # Local data persistence
├── chrome-extension/       # 👁️ The Eyes
│   ├── manifest.json       # Extension config
│   ├── background.js       # Event capture & API communication
│   └── popup.js            # User interface
├── hackathon-challenge/    # 🧪 The Lab
│   ├── main.py             # Hybrid routing logic (The Core Challenge)
│   └── benchmark.py        # Performance evaluation suite
└── docs/                   # 📚 Documentation

System Architecture

Nova operates on a three-tier architecture designed for minimal latency and maximum privacy:

1. Presentation Layer: The Chrome Extension captures user interactions (page loads, inputs) and sends them to the local server.
2. Application Layer: The Monitor Backend receives events, manages session state, and orchestrates the AI analysis.
3. Intelligence Layer: The Hybrid Router determines the best execution path:
   • Local Path: Uses FunctionGemma via Cactus for fast, private inference (typical latency: 50-100ms).
   • Cloud Path: Falls back to Gemini 3.0 Flash when the local model's confidence is low or the task is too complex (typical latency: 500ms+).

Our Approach

The core innovation in Nova is its Hybrid Routing Strategy. Instead of relying solely on the cloud (slow, privacy-invasive) or solely on the edge (limited reasoning), we implemented a dynamic router:

1. Confidence-Based Routing: We utilize the confidence scores returned by the Cactus engine. If the local model is >99% confident, we use its result immediately.
2. Latency Optimization: By prioritizing the local model, we achieve near-instantaneous feedback for the user, essential for a real-time monitoring tool.
3. Graceful Fallback: If the local model struggles or returns a low-confidence score, the system seamlessly delegates the task to Gemini Cloud, ensuring high accuracy even for ambiguous inputs.

Getting Started

Prerequisites

• macOS with Apple Silicon (M1/M2/M3/M4).
• Python 3.9+.
• Google Chrome.

Installation

1. Clone the repository:

   git clone https://github.com/switchchat/extension.git
   cd extension

2. Run the setup script: This will create a virtual environment, install dependencies, and start the backend.

   ./start.sh

3. Install the Chrome Extension:

   • Go to chrome://extensions.
   • Enable "Developer mode".
   • Click "Load unpacked" and select the chrome-extension folder.

4. Start Browsing:

   • Open the extension popup to see the connection status.
   • Browse the web, and watch Nova analyze your activity in real-time!

License

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

---

Built with 🌵 Cactus and 🧠 Google DeepMind