The author, an Android developer with 25 years of experience, developed an AI-agent application over six months. The project involved 1,600 commits and was built using AI-assisted coding tools, transitioning from Gemini CLI to Claude Code at the 276-commit mark.
The application runs models locally on the device using the LiteRT-LM engine, a wrapper for TensorFlow Lite. It supports Gemma 4 E4B and E2B models, requiring approximately 3.4 GB for the model file and at least 2 GB of free memory. Cloud-based models and external tools are optional and require manual API key or server configuration.
The architecture evolved into a graph-based system with 14 node types to manage complex tasks. The developer implemented validation gates for error handling, context management for individual nodes, and a composition system to reuse pipelines. The system prevents unauthorized actions by requiring user confirmation for sensitive operations.
The project faced significant challenges with Android AppFunctions, as the necessary EXECUTE_APP_FUNCTIONS permission is restricted to system-level applications. Consequently, the developer shifted to the Model Context Protocol (MCP) for tool integration. The application currently supports Android 14 and above.
