// the find
NickMonrad/kernel-ai-assistant
Jandal AI — Local-first Android AI assistant with on-device LLM inference, semantic memory, and extensible skill framework
Jandal AI is a local-first Android assistant that runs Gemma-4 on-device via LiteRT-LM, with sqlite-vec for semantic memory and a native Kotlin skill layer for deterministic device actions. It targets Android users who want an always-on assistant without their data leaving the phone. Pre-release, not yet on the Play Store.
The Brain-Memory-Action split is a clean architecture decision — the QuickIntentRouter handles deterministic fast paths (set alarm, start timer) without hitting the LLM at all, which is the right call. The STT stack is genuinely flexible: Vosk, Sherpa-ONNX Zipformer, SenseVoice, Whisper tiny.en, and Paraformer are all wired up, so you can match the engine to the device. Using sqlite-vec for on-device vector storage avoids dragging in a separate vector database. Privacy model is genuine: chat, memories, and inference are all local by default, with external calls limited to explicitly web-backed skills.
The required model download is ~2.4 GB before the app is usable — that is a serious activation problem on metered connections or budget devices, and there is no apparent on-device quantisation path to a smaller tier. The repo is classified as 'C' on GitHub but is entirely Kotlin; whoever set the language tag did not notice. The `.squad/` directory contains an elaborate multi-agent AI orchestration framework (agent personas, casting templates with Futurama character assignments, ceremonies, routing configs) that the developer is using to manage their own codebase — interesting as a workflow experiment, but it adds hundreds of files of overhead that have nothing to do with the app. Wasm skill extensibility, which is the main story for third-party developers, is listed as a research path with no timeline.