Name	Name	Last commit message	Last commit date
parent directory ..
app	app
gradle/wrapper	gradle/wrapper
.editorconfig	.editorconfig
.gitignore	.gitignore
README.md	README.md
build.gradle.kts	build.gradle.kts
detekt.yml	detekt.yml
gradle.properties	gradle.properties
gradlew	gradlew
gradlew.bat	gradlew.bat
settings.gradle.kts	settings.gradle.kts

RunAnywhere AI - Android Example

A production-ready reference app demonstrating the RunAnywhere Kotlin SDK capabilities for on-device AI. This app showcases how to build privacy-first, offline-capable AI features with LLM chat, speech-to-text, text-to-speech, and a complete voice assistant pipeline—all running locally on your device.

🚀 Running This App (Local Development)

Important: This sample app consumes the RunAnywhere Kotlin SDK as a local Gradle included build. Before opening this project, you must first build the SDK's native libraries.

First-Time Setup

# 1. Navigate to the Kotlin SDK directory
cd runanywhere-sdks/sdk/runanywhere-kotlin

# 2. Run the setup script (~10-15 minutes on first run)
#    This builds the native C++ JNI libraries and sets testLocal=true
./scripts/build-kotlin.sh --setup

# 3. Open this sample app in Android Studio
#    File > Open > examples/android/RunAnywhereAI

# 4. Wait for Gradle sync to complete

# 5. Connect an Android device (ARM64 recommended) or use an emulator

# 6. Click Run

How It Works

This sample app uses settings.gradle.kts with includeBuild() to reference the local Kotlin SDK:

This Sample App → Local Kotlin SDK (sdk/runanywhere-kotlin/)
                          ↓
              Local JNI Libraries (sdk/runanywhere-kotlin/src/androidMain/jniLibs/)
                          ↑
           Built by: ./scripts/build-kotlin.sh --setup

The build-kotlin.sh --setup script:

Downloads dependencies (Sherpa-ONNX, ~500MB)
Builds the native C++ libraries from runanywhere-commons
Copies JNI .so files to sdk/runanywhere-kotlin/src/androidMain/jniLibs/
Sets runanywhere.testLocal=true in gradle.properties

After Modifying the SDK

Kotlin SDK code changes: Rebuild in Android Studio or run ./gradlew assembleDebug

C++ code changes (in runanywhere-commons):

cd sdk/runanywhere-kotlin
./scripts/build-kotlin.sh --local --rebuild-commons

Try It Now

Download the app from Google Play Store to try it out.

Screenshots

Features

This sample app demonstrates the full power of the RunAnywhere SDK:

Feature	Description	SDK Integration
AI Chat	Interactive LLM conversations with streaming responses	`RunAnywhere.generateStream()`
Thinking Mode	Support for models with `<think>...</think>` reasoning	Thinking tag parsing
Real-time Analytics	Token speed, generation time, inference metrics	`MessageAnalytics`
Speech-to-Text	Voice transcription with batch & live modes	`RunAnywhere.transcribe()`
Text-to-Speech	Neural voice synthesis with Piper TTS	`RunAnywhere.synthesize()`
Voice Assistant	Full STT -> LLM -> TTS pipeline with auto-detection	`RunAnywhere.processVoice()`
Model Management	Download, load, and manage multiple AI models	`RunAnywhere.downloadModel()`
Storage Management	View storage usage and delete models	`RunAnywhere.storageInfo()`
Offline Support	All features work without internet	On-device inference

Architecture

The app follows modern Android architecture patterns:

┌─────────────────────────────────────────────────────────────────┐
│                      Jetpack Compose UI                          │
│  ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌────────┐ │
│  │  Chat    │ │   STT    │ │   TTS    │ │  Voice   │ │Settings│ │
│  │  Screen  │ │  Screen  │ │  Screen  │ │  Screen  │ │ Screen │ │
│  └────┬─────┘ └────┬─────┘ └────┬─────┘ └────┬─────┘ └───┬────┘ │
├───────┼────────────┼────────────┼────────────┼───────────┼──────┤
│       ▼            ▼            ▼            ▼           ▼      │
│  ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌────────┐ │
│  │  Chat    │ │   STT    │ │   TTS    │ │  Voice   │ │Settings│ │
│  │ViewModel │ │ViewModel │ │ViewModel │ │ViewModel │ │ViewModel│
│  └────┬─────┘ └────┬─────┘ └────┬─────┘ └────┬─────┘ └───┬────┘ │
├───────┴────────────┴────────────┴────────────┴───────────┴──────┤
│                                                                  │
│                    RunAnywhere Kotlin SDK                        │
│  ┌──────────────────────────────────────────────────────────┐   │
│  │  Extension Functions (generate, transcribe, synthesize)   │   │
│  │  EventBus (LLMEvent, STTEvent, TTSEvent, ModelEvent)     │   │
│  │  Model Management (download, load, unload, delete)        │   │
│  └──────────────────────────────────────────────────────────┘   │
│                              │                                   │
│           ┌──────────────────┴──────────────────┐               │
│           ▼                                      ▼               │
│  ┌─────────────────┐                  ┌─────────────────┐       │
│  │   LlamaCpp      │                  │   ONNX Runtime  │       │
│  │   (LLM/GGUF)    │                  │   (STT/TTS)     │       │
│  └─────────────────┘                  └─────────────────┘       │
└─────────────────────────────────────────────────────────────────┘

Key Architecture Decisions

MVVM Pattern — ViewModels manage UI state with StateFlow, Compose observes changes
Single Activity — Jetpack Navigation Compose handles all screen transitions
Coroutines & Flow — All async operations use Kotlin coroutines with structured concurrency
EventBus Pattern — SDK events (model loading, generation, etc.) propagate via EventBus.events
Repository Abstraction — ConversationStore persists chat history

Project Structure

RunAnywhereAI/
├── app/
│   ├── src/main/
│   │   ├── java/com/runanywhere/runanywhereai/
│   │   │   ├── RunAnywhereApplication.kt      # SDK initialization, model registration
│   │   │   ├── MainActivity.kt                # Entry point, initialization state handling
│   │   │   │
│   │   │   ├── data/
│   │   │   │   └── ConversationStore.kt       # Chat history persistence
│   │   │   │
│   │   │   ├── domain/
│   │   │   │   ├── models/
│   │   │   │   │   ├── ChatMessage.kt         # Message data model with analytics
│   │   │   │   │   └── SessionState.kt        # Voice session states
│   │   │   │   └── services/
│   │   │   │       └── AudioCaptureService.kt # Microphone audio capture
│   │   │   │
│   │   │   ├── presentation/
│   │   │   │   ├── chat/
│   │   │   │   │   ├── ChatScreen.kt          # LLM chat UI with streaming
│   │   │   │   │   ├── ChatViewModel.kt       # Chat logic, thinking mode
│   │   │   │   │   └── components/
│   │   │   │   │       └── MessageInput.kt    # Chat input component
│   │   │   │   │
│   │   │   │   ├── stt/
│   │   │   │   │   ├── SpeechToTextScreen.kt  # STT UI with waveform
│   │   │   │   │   └── SpeechToTextViewModel.kt # Batch & live transcription
│   │   │   │   │
│   │   │   │   ├── tts/
│   │   │   │   │   ├── TextToSpeechScreen.kt  # TTS UI with playback
│   │   │   │   │   └── TextToSpeechViewModel.kt # Synthesis & audio playback
│   │   │   │   │
│   │   │   │   ├── voice/
│   │   │   │   │   ├── VoiceAssistantScreen.kt # Full voice pipeline UI
│   │   │   │   │   └── VoiceAssistantViewModel.kt # STT→LLM→TTS orchestration
│   │   │   │   │
│   │   │   │   ├── settings/
│   │   │   │   │   ├── SettingsScreen.kt      # Storage & model management
│   │   │   │   │   └── SettingsViewModel.kt   # Storage info, cache clearing
│   │   │   │   │
│   │   │   │   ├── models/
│   │   │   │   │   ├── ModelSelectionBottomSheet.kt # Model picker UI
│   │   │   │   │   └── ModelSelectionViewModel.kt   # Download & load logic
│   │   │   │   │
│   │   │   │   ├── navigation/
│   │   │   │   │   └── AppNavigation.kt       # Bottom nav, routing
│   │   │   │   │
│   │   │   │   └── common/
│   │   │   │       └── InitializationViews.kt # Loading/error states
│   │   │   │
│   │   │   └── ui/theme/
│   │   │       ├── Theme.kt                   # Material 3 theming
│   │   │       ├── AppColors.kt               # Color palette
│   │   │       ├── Type.kt                    # Typography
│   │   │       └── Dimensions.kt              # Spacing constants
│   │   │
│   │   ├── res/                               # Resources (icons, strings)
│   │   └── AndroidManifest.xml                # Permissions, app config
│   │
│   ├── src/test/                              # Unit tests
│   └── src/androidTest/                       # Instrumentation tests
│
├── build.gradle.kts                           # Project build config
├── settings.gradle.kts                        # Module settings
└── README.md                                  # This file

Quick Start

Prerequisites

Android Studio Hedgehog (2023.1.1) or later
Android SDK 24+ (Android 7.0 Nougat)
JDK 17+
Device/Emulator with arm64-v8a architecture (recommended: physical device)
~2GB free storage for AI models

Clone & Build

# Clone the repository
git clone https://github.com/RunanywhereAI/runanywhere-sdks.git
cd runanywhere-sdks/examples/android/RunAnywhereAI

# Build debug APK
./gradlew assembleDebug

# Install on connected device
./gradlew installDebug

Run via Android Studio

Open the project in Android Studio
Wait for Gradle sync to complete
Select a physical device (arm64 recommended) or emulator
Click Run or press Shift + F10

Run via Command Line

# Install and launch
./gradlew installDebug
adb shell am start -n com.runanywhere.runanywhereai.debug/.MainActivity

SDK Integration Examples

Initialize the SDK

The SDK is initialized in RunAnywhereApplication.kt:

// Initialize SDK with development environment
RunAnywhere.initialize(environment = SDKEnvironment.DEVELOPMENT)

// Complete services initialization (device registration)
RunAnywhere.completeServicesInitialization()

// Register AI backends
LlamaCPP.register(priority = 100)  // LLM backend (GGUF models)
ONNX.register(priority = 100)      // STT/TTS backend

// Register models
RunAnywhere.registerModel(
    id = "smollm2-360m-q8_0",
    name = "SmolLM2 360M Q8_0",
    url = "https://huggingface.co/prithivMLmods/SmolLM2-360M-GGUF/...",
    framework = InferenceFramework.LLAMA_CPP,
    memoryRequirement = 500_000_000,
)

Download & Load a Model

// Download with progress tracking
RunAnywhere.downloadModel("smollm2-360m-q8_0").collect { progress ->
    println("Download: ${(progress.progress * 100).toInt()}%")
}

// Load into memory
RunAnywhere.loadLLMModel("smollm2-360m-q8_0")

Stream Text Generation

// Generate with streaming
RunAnywhere.generateStream(prompt).collect { token ->
    // Display token in real-time
    displayToken(token)
}

// Or non-streaming
val result = RunAnywhere.generate(prompt)
println("Response: ${result.text}")

Speech-to-Text

// Load STT model
RunAnywhere.loadSTTModel("sherpa-onnx-whisper-tiny.en")

// Transcribe audio bytes
val transcription = RunAnywhere.transcribe(audioBytes)
println("Transcription: $transcription")

Text-to-Speech

// Load TTS voice
RunAnywhere.loadTTSVoice("vits-piper-en_US-lessac-medium")

// Synthesize speech
val result = RunAnywhere.synthesize(text, TTSOptions(
    rate = 1.0f,
    pitch = 1.0f,
))
// result.audioData contains WAV audio bytes

Voice Pipeline (STT → LLM → TTS)

// Process voice through full pipeline
val result = RunAnywhere.processVoice(audioData)

if (result.speechDetected) {
    println("User said: ${result.transcription}")
    println("AI response: ${result.response}")
    // result.synthesizedAudio contains TTS audio
}

Key Screens Explained

1. Chat Screen (`ChatScreen.kt`)

What it demonstrates:

Streaming text generation with real-time token display
Thinking mode support (<think>...</think> tags)
Message analytics (tokens/sec, time to first token)
Conversation history management
Model selection bottom sheet integration

Key SDK APIs:

RunAnywhere.generateStream() — Streaming generation
RunAnywhere.generate() — Non-streaming generation
RunAnywhere.cancelGeneration() — Stop generation
EventBus.events.filterIsInstance<LLMEvent>() — Listen for LLM events

2. Speech-to-Text Screen (`SpeechToTextScreen.kt`)

What it demonstrates:

Batch mode: Record full audio, then transcribe
Live mode: Real-time streaming transcription
Audio level visualization
Transcription metrics (confidence, RTF, word count)

Key SDK APIs:

RunAnywhere.loadSTTModel() — Load Whisper model
RunAnywhere.transcribe() — Batch transcription
RunAnywhere.transcribeStream() — Streaming transcription

3. Text-to-Speech Screen (`TextToSpeechScreen.kt`)

What it demonstrates:

Neural voice synthesis with Piper TTS
Speed and pitch controls
Audio playback with progress
Fun sample texts for testing

Key SDK APIs:

RunAnywhere.loadTTSVoice() — Load TTS model
RunAnywhere.synthesize() — Generate speech audio
RunAnywhere.stopSynthesis() — Cancel synthesis

4. Voice Assistant Screen (`VoiceAssistantScreen.kt`)

What it demonstrates:

Complete voice AI pipeline
Automatic speech detection with silence timeout
Continuous conversation mode
Model status tracking for all 3 components (STT, LLM, TTS)

Key SDK APIs:

RunAnywhere.startVoiceSession() — Start voice session
RunAnywhere.processVoice() — Process audio through pipeline
RunAnywhere.voiceAgentComponentStates() — Check component status

5. Settings Screen (`SettingsScreen.kt`)

What it demonstrates:

Storage usage overview
Downloaded model management
Model deletion with confirmation
Cache clearing

Key SDK APIs:

RunAnywhere.storageInfo() — Get storage details
RunAnywhere.deleteModel() — Remove downloaded model
RunAnywhere.clearCache() — Clear temporary files

Testing

Run Unit Tests

./gradlew test

Run Instrumentation Tests

./gradlew connectedAndroidTest

Run Lint & Static Analysis

# Detekt static analysis
./gradlew detekt

# ktlint formatting check
./gradlew ktlintCheck

# Android lint
./gradlew lint

Debugging

Enable Verbose Logging

Filter logcat for RunAnywhere SDK logs:

adb logcat -s "RunAnywhere:D" "RunAnywhereApp:D" "ChatViewModel:D"

Common Log Tags

Tag	Description
`RunAnywhereApp`	SDK initialization, model registration
`ChatViewModel`	LLM generation, streaming
`STTViewModel`	Speech transcription
`TTSViewModel`	Speech synthesis
`VoiceAssistantVM`	Voice pipeline
`ModelSelectionVM`	Model downloads, loading

Memory Profiling

Open Android Studio Profiler
Select your app process
Record memory allocations during model loading
Expected: ~300MB-2GB depending on model size

Configuration

Build Variants

Variant	Description
`debug`	Development build with debugging enabled
`release`	Optimized build with R8/ProGuard
`benchmark`	Release-like build for performance testing

Environment Variables (for release builds)

export KEYSTORE_PATH=/path/to/keystore.jks
export KEYSTORE_PASSWORD=your_password
export KEY_ALIAS=your_alias
export KEY_PASSWORD=your_key_password

Supported Models

LLM Models (LlamaCpp/GGUF)

Model	Size	Memory	Description
SmolLM2 360M Q8_0	~400MB	500MB	Fast, lightweight chat
Qwen 2.5 0.5B Q6_K	~500MB	600MB	Multilingual, efficient
LFM2 350M Q4_K_M	~200MB	250MB	LiquidAI, ultra-compact
Llama 2 7B Chat Q4_K_M	~4GB	4GB	Powerful, larger model
Mistral 7B Instruct Q4_K_M	~4GB	4GB	High quality responses

STT Models (ONNX/Whisper)

Model	Size	Description
Sherpa Whisper Tiny (EN)	~75MB	English transcription

TTS Models (ONNX/Piper)

Model	Size	Description
Piper US English (Medium)	~65MB	Natural American voice
Piper British English (Medium)	~65MB	British accent

Known Limitations

ARM64 Only — Native libraries built for arm64-v8a only (x86 emulators not supported)
Memory Usage — Large models (7B+) require devices with 6GB+ RAM
First Load — Initial model loading takes 1-3 seconds (cached afterward)
Thermal Throttling — Extended inference may trigger device throttling on some devices

Contributing

See CONTRIBUTING.md for guidelines.

Development Setup

# Fork and clone
git clone https://github.com/YOUR_USERNAME/runanywhere-sdks.git
cd runanywhere-sdks/examples/android/RunAnywhereAI

# Create feature branch
git checkout -b feature/your-feature

# Make changes and test
./gradlew assembleDebug
./gradlew test
./gradlew detekt ktlintCheck

# Commit and push
git commit -m "feat: your feature description"
git push origin feature/your-feature

# Open Pull Request

License

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

Support

Discord: Join our community
GitHub Issues: Report bugs
Email: san@runanywhere.ai
Twitter: @RunanywhereAI

FilesExpand file tree

RunAnywhereAI

Directory actions

More options

Directory actions

More options

Latest commit

History

RunAnywhereAI

Folders and files

parent directory

README.md

RunAnywhere AI - Android Example

🚀 Running This App (Local Development)

First-Time Setup

How It Works

After Modifying the SDK

Try It Now

Screenshots

Features

Architecture

Key Architecture Decisions

Project Structure

Quick Start

Prerequisites

Clone & Build

Run via Android Studio

Run via Command Line

SDK Integration Examples

Initialize the SDK

Download & Load a Model

Stream Text Generation

Speech-to-Text

Text-to-Speech

Voice Pipeline (STT → LLM → TTS)

Key Screens Explained

1. Chat Screen (ChatScreen.kt)

2. Speech-to-Text Screen (SpeechToTextScreen.kt)

3. Text-to-Speech Screen (TextToSpeechScreen.kt)

4. Voice Assistant Screen (VoiceAssistantScreen.kt)

5. Settings Screen (SettingsScreen.kt)

Testing

Run Unit Tests

Run Instrumentation Tests

Run Lint & Static Analysis

Debugging

Enable Verbose Logging

Common Log Tags

Memory Profiling

Configuration

Build Variants

Environment Variables (for release builds)

Supported Models

LLM Models (LlamaCpp/GGUF)

STT Models (ONNX/Whisper)

TTS Models (ONNX/Piper)

Known Limitations

Contributing

Development Setup

License

Support

Related Documentation

1. Chat Screen (`ChatScreen.kt`)

2. Speech-to-Text Screen (`SpeechToTextScreen.kt`)

3. Text-to-Speech Screen (`TextToSpeechScreen.kt`)

4. Voice Assistant Screen (`VoiceAssistantScreen.kt`)

5. Settings Screen (`SettingsScreen.kt`)