I'm a Lead AI Engineer working where interpretability, reasoning, and efficient inference meet. I like taking models apart to see how they actually think (logit lenses, steering vectors, attention maps), then turning that understanding into systems that reason harder and run faster.

In practice that means symbolic reasoning on the ARC-AGI benchmark, self-trained speculative-decoding heads squeezed onto Apple Silicon, and applied ML across medical imaging and beyond. I build from first principles and report results honestly: losses reported as losses, artifacts flagged as artifacts.

Currently exploring: interpretability-guided speculative decoding and SAE-based activation steering.

_{Logit lens: watching “Paris” win the prediction as the thought forms, layer by layer, inside Qwen3-0.6B.}

A deterministic 2D physics fighting engine where small local LLMs (served on-device via MLX) drive stickmen through a hybrid loop: the model picks a tactical intent, a 60 Hz executor turns it into moves. Plus a self-play ELO league where losing agents rewrite their own policies via LLM reflection. TypeScript monorepo, 85 tests, frame-exact replays.

_{A recorded match, replayed frame-exact. Each fighter is driven by a small local LLM; the bubble is the tactical intent it just picked.}

• Mechanistic interpretability: logit lens, steering directions, attention analysis; reading a model mid-thought
• Reasoning: symbolic and neural approaches to ARC-AGI, verification-first program search
• Efficient inference: self-trained EAGLE / speculative decoding, batching, running LLMs on Apple Silicon (MLX)
• Applied ML: medical imaging, anomaly detection, and the occasional quantum circuit