On-Device AI Coach with Expo

An on-device AI wellness coach built with Expo and @react-native-ai/mlc, running Llama-3.2-3B-Instruct model entirely on your device.

Features

• 🤖 Real On-Device AI: Uses MLC (Machine Learning Compilation) to run Llama-3.2-3B-Instruct locally
• 🔒 Privacy-First: Your data never leaves the device
• ⚡ Fast Responses: Local inference means instant replies
• 📱 Native Mobile: Works on iOS and Android (requires Expo Dev Build)
• 💬 Wellness Coach: Get personalized advice for wellness, productivity, and mental health

Setup

Prerequisites

• Node.js and npm installed
• Expo CLI installed (npm install -g expo-cli)
• For iOS: Xcode and CocoaPods
• For Android: Android Studio and JDK

Installation

1. Install dependencies

   npm install

2. Prebuild native directories (already done, but you can regenerate)

   npx expo prebuild --clean

3. For iOS (fix CocoaPods encoding if needed):

   export LANG=en_US.UTF-8
   cd ios && pod install && cd ..

4. Start the app

   npx expo start --dev-client

5. Run on device

   • Android: npx expo run:android
   • iOS: npx expo run:ios

Important Notes

First Run

On first launch, the model will be downloaded and initialized. This may take:

• Download: 10-30 minutes depending on connection (~3GB)
• Preparation: 1-2 minutes
• Subsequent runs: Much faster (model is cached)

Model Information

• Model: Llama-3.2-3B-Instruct
• Size: ~3GB (downloaded on first use)
• Provider: MLC (Machine Learning Compilation)
• Inference: On-device, no API calls

Requirements

• React Native New Architecture: Enabled (already configured)
• Increased Memory Limit: Configured via MLC plugin
• Storage: ~3GB free space needed
• Internet: Required for initial model download only

Project Structure

thrive/
├── ai/
│   └── ai.ts              # Real MLC on-device AI engine
├── components/
│   ├── CoachInput.tsx     # Input component
│   └── CoachOutput.tsx    # Output component
├── app/
│   └── (tabs)/
│       └── index.tsx       # Main AI coach screen
└── ios/                    # Native iOS (generated by prebuild)
└── android/                # Native Android (generated by prebuild)

Development

Using Expo Dev Client

Since this uses native modules, you need Expo Dev Client (not Expo Go):

# Build dev client
npx expo run:android
# or
npx expo run:ios

# Then start development server
npx expo start --dev-client

Troubleshooting

Model won't download:

• Check internet connection
• Ensure ~3GB free storage
• Check device permissions

Build errors:

• Run npx expo prebuild --clean again
• For iOS: Fix CocoaPods encoding: export LANG=en_US.UTF-8
• Clear cache: npx expo start -c

Memory issues:

• The MLC plugin already configures increased memory limit
• Close other apps if needed

Learn more

• Expo documentation
• React Native AI documentation
• MLC documentation
• Llama 3.2 model information