A repository for exploring on-device agentic AI on Android — running offline using small language models (SLMs), tool calling, and agentic frameworks.
This repo contains two apps built on a shared multi-module core:
- EdgeLab — a test and validation platform for on-device LLM tool calling
- CyclingCopilot — an AI cycling assistant demo (in progress)
Built with LiteRT-LM, JetBrains Koog, and HuggingFace.
| Resource | Description |
|---|---|
| 🌐 edgeagentlab.dev | Microsite with test visualizations and project overview |
| Part #1 — From Flat Notifications to Edge AI | Initial concept and motivation |
| Part #2 — Function Calling with Edge AI | Deep dive into on-device tool calling challenges |
| Part #3 — Let's talk about FunctionGemma | Exploring FunctionGemma and where it fits in mobile development |
The project uses a multi-module KMP-ready architecture with a shared Android core:
| Module | Type | Description |
|---|---|---|
:data |
Pure Kotlin | Data models, provider interfaces |
:agent |
Pure Kotlin | Agent logic, tool handlers, inference interfaces |
:presentation |
Pure Kotlin | MVI ViewModels, test engine (KMP-ready) |
:core |
Android library | Shared infrastructure: inference, download, OAuth, storage |
:app:explorer |
Android app | EdgeLab — model testing and tool calling validation |
:app:copilot |
Android app | CyclingCopilot — on-device AI cycling assistant |
Module dependencies: :data ← :agent ← :presentation ← :core ← :app:explorer, :app:copilot
For full technical details see docs/architecture.md.
| Document | Description |
|---|---|
docs/architecture.md |
Full technical architecture and implementation details |
docs/edgelab/roadmap.md |
EdgeLab roadmap |
docs/cyclingcopilot/roadmap.md |
CyclingCopilot roadmap |
docs/cyclingcopilot/ui-architecture.md |
CyclingCopilot UI design and screen specs |
- Kotlin (Multiplatform-ready modules)
- Jetpack Compose
- LiteRT-LM
- JetBrains Koog
- HuggingFace Hub integration
- Firebase Crashlytics
This is a prototype built for fast iteration, experimentation, and learning — exploring the limits of edge AI and agentic workflows on Android.
MIT License — see the LICENSE file for details.