Picks up every call. Takes orders, makes reservations, answers questions. 60-second menu onboarding from a photo via Gemini Pro Vision, natural voice via ElevenLabs, live dashboard via Supabase Realtime. Built solo in 7 hours.

Taught a robot arm to pour latte art by fine-tuning Physical Intelligence's π0 vision-language-action model on demonstration data. Runs on an OpenDroid R2D3 arm. Open source.

Python library for multi-agent systems. Agents store structured claims with evidence and provenance, query shared memory before doing redundant work, and get automatic contradiction alerts when findings conflict. FAISS + SQLite backend. 56% reuse rate, 17.5% token reduction on a 3-agent benchmark.

Natural language to grasp-and-place in a cluttered MuJoCo scene. Gemini handles grounding and target selection, MuJoCo handles physics.

Trains, evaluates, and compares multiple models end-to-end with a retraining loop that improves baseline accuracy by up to 12%. ~90% less manual tuning on the datasets we ran it on.

Serves JSON files when present, otherwise proxies to the real API. Parameterized mocks with latency injection for frontend dev loops. Published on npm.

Binary classifier for region of attraction estimation. Given a state, predicts whether the system will reach the goal or fail. Part of ongoing robotics ML research at PRACSYS Lab, Rutgers.