GemYum - On-Device AI Nutrition Tracking

🏆 Google - Gemma 3n Hackathon

GemYum is a privacy-first Android application that uses Google's Gemma 3n model for on-device food recognition and nutrition tracking. Built for the Google - Gemma 3n Hackathon, it demonstrates practical edge AI applications in health and wellness.

✨ Key Features

• 🤖 On-Device AI: Runs Gemma 3n model locally - no cloud required
• 📸 Instant Food Recognition: Point camera at food for instant nutrition data
• 📊 Comprehensive Database: 700+ foods with glycemic index data
• 🔒 Privacy First: All processing happens on your device
• 📱 Offline Capable: Works without internet connection
• 💚 Health Connect Integration: Syncs with Android Health platform

📦 Downloads

• Latest Release (v1.4.0) - Standard APK (requires internet for models)
• Offline Package - Pre-bundled with AI models
• All Releases - Version history and changelogs

🚀 Quick Start

Prerequisites

• Android device (API 24+) with 6GB+ RAM
• 5GB free storage for AI models
• WiFi/Internet connection for initial model download
• Android Studio (for building from source)

Installation

Option 1: Download Pre-built APK (Recommended)

1. Download GemYum v1.4.0

   • Download APK from Google Drive
   • Enable "Install from Unknown Sources" in Android settings
   • Install the APK on your device

2. Download AI Models (Required)

   • Launch the app after installation
   • Connect to WiFi/Internet for model download
   • The app will download Gemma 3n models (~3-4GB):
     • E2B Model - 2GB (Required)
     • E4B Model - 4GB (Optional, better accuracy)
   • Models are downloaded once and work offline thereafter

Option 2: Offline Installation (No Internet Required)

For devices without internet access, use the pre-bundled APK with models:

• Offline Installation Package
• Contains APK with embedded AI models (larger file size ~6GB)
• No internet connection required after installation

Building from Source

# Clone repository
git clone https://github.com/Stellife/GemYum.git
cd GemYum

# Build APK
./gradlew assembleRelease

# Install on device
adb install app/build/outputs/apk/release/app-release.apk

📱 Usage

1. Setup: Download AI models on first launch
2. Camera Mode: Point at food for instant recognition
3. Search: Type food names for nutrition lookup
4. Track: View daily nutrition summaries
5. Health Sync: Export to Health Connect

🏗️ Technical Architecture

• AI Model: Gemma 3n (3B parameters) quantized for mobile
• Framework: TensorFlow Lite with GPU acceleration
• Database: SQLite with 700+ foods and glycemic index
• UI: Jetpack Compose with Material 3
• Platform: Android (Kotlin)

📊 Model Performance

• Inference Time: ~2-3 seconds per image
• Accuracy: 85%+ on common foods
• Memory Usage: ~4GB during inference
• Battery Impact: Minimal with GPU acceleration

🤝 Contributing

Contributions are welcome! Please read our Contributing Guide for details.

📄 License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

🙏 Acknowledgments

• Google for the Gemma model and TensorFlow Lite
• Kaggle for hosting the competition
• The Android ML community for inspiration

📧 Contact

For questions about this Kaggle submission, please open an issue on GitHub.

---

Competition Entry: This project is submitted to the Google - Gemma 3n Hackathon.