On-device AI for hardware people actually own.

I build local-first systems on Apple Silicon and Android: voice agents, small-model training and fine-tunes, chess engines, and practical desktop/mobile tools. The throughline is simple — privacy by default, measurable performance, and evidence before hype.

• Current focus: offline voice pipelines, chess evaluation / self-play, Android readers & on-device apps, vectorization tooling
• Working rule: if a result matters, it should have a config, a log, a benchmark, or a release artifact
• Primary targets: macOS (Metal / MLX / MPS), Android (Kotlin / on-device inference)

MacBot · Python · ⭐ flagship voice stack
Offline macOS voice assistant with an interruptible pipeline: VAD → Whisper STT → local LLM (llama.cpp) → neural TTS. Metal acceleration, native macOS tools, secure local RAG, and a realtime WebSocket dashboard.

GemmaFischer · Python · HF adapters
Fine-tuned Gemma-3 as a UCI chess engine + interactive tutor. LoRA experts (UCI / tutor / director), MoE-style routing, and Apple Silicon MPS training/inference.

Matrix0 · Python
AlphaZero-style chess engine with SSL heads for tactical pattern recognition (threats, pins, forks, control). Multi-task RL on a ResNet backbone, MCTS, MPS optimization, WebUI, and fixed-baseline benchmarking.

svg-X · TypeScript · releases
Desktop + CLI image→SVG converter (Electron/React). Triple pipeline: Potrace B&W, color posterization with Bezier fitting, and centerline stroke mode. Exports SVG / EPS / DXF / JSON paths.

XReader · Kotlin
Native Android e-reader for DRM-free personal libraries. SAF import, app-private storage, multi-format conversion into a Readium path, local notes/bookmarks/search — no cloud required.

ChessTrainer · Java
On-device Android chess trainer with local Stockfish / LC0 engines, lessons, analysis, and foldable-friendly UI.

Breakout- · Kotlin
Offline brick-breaker built for phones and foldables: OpenGL rendering, fixed-step physics, multiple modes and powerups, no network or analytics SDK.

Work in progress tends to stay private until configs, evals, and docs are honest enough to ship. Themes that show up across public and private repos:

• Local LLMs — post-training, adapters, and small-model stacks aimed at real devices
• Realtime voice — interruptible STT → LLM → TTS loops with latency budgets
• Chess + evaluation — engines, tutors, self-play, and reproducible match gates
• Android privacy tools — readers, on-device assistants, offline-first mobile apps

• GitHub: github.com/lukifer23
• Models: Hugging Face (GemmaFischer LoRA collection)

For project-specific questions, open an issue in the relevant repo. For collaboration, start with the project closest to what you want to build or test.