I get big AI models running on small hardware, and I build AI agents that verify their own work.
- Edge & local LLM inference — llama.cpp on NPUs, MoE models on constrained RAM, Raspberry Pi
- AI agent engineering — tool loops, retrieval, persistent memory
- Research → working code — reimplementing papers with held-out, honestly-reported evaluations
- Game modding & dev tooling — BepInEx, Godot, MCP
| Project | What it does |
|---|---|
| zero-mem-pi | Zero-token long-term memory for the pi coding agent — from-scratch reimplementation of the Zero-Mem paper (arXiv:2607.29377). Hybrid BM25 + dense + graph retrieval, int8-quantized store, validated on LoCoMo10 with significance testing. |
| Axera-AX8850-GGUF-Support | llama.cpp backend running GGUF models directly on the AX8850 NPU — 24–30 t/s decode on a Raspberry Pi 5. |
| GREYHACKLLM | An autonomous coding agent that operates the simulated computer inside the game Grey Hack — BepInEx C# hook, file/TCP bridge, GLM-powered tool loop with verify-before-done discipline. |
| 3d-model-generator | Local text/image → rigged, animated, textured low-poly GLB models (TRELLIS.2 + FLUX). |
| semopt | LLM-aided compiler optimization — validated, contract-preserving source rewrites (SeGaBench protocol). |
| llmwebchat | Provider-agnostic, local-first LLM chat UI with a secrets-holding proxy. |
Deep in local LLM inference, retrieval evaluation, and agent memory. Open to freelance and contract work.
Python · TypeScript · C/C++ · C# · Kotlin