A mix of research, infrastructure, and product work. Most of it goes from paper to running code to real users or real hardware.
Evaluation platform for embodied / VLA manipulation policies. VLM-as-Judge scoring with dual-judge agreement, typed metrics, and a live dashboard.
A RoboEval-native simulation benchmark for warehouse packing: manipulation policies scored on partial-completion across tasks and robots in Isaac.
Fine-tuned a SmolVLA vision-language-action policy and deployed it to the SO-101 arm, with mechanistic-interpretability probes into what the policy learns.
Biologically-inspired locomotion controller: simulates the full 13,101-neuron Drosophila ventral nerve cord in Brian2 + FlyGym, then deploys to a physical hexapod.
LLM-native crypto wallet with a 4-check safety engine on Base. Natural-language intents across send, swap, batch, chain, streaming, and bridging.
Local-first prediction market forecasting agent with LLM-driven sub-question decomposition, devil's advocate debate, and automated Kalshi trading.
Data-driven marketplace for physical ads. Zillow-style discovery on a map, top 6 at YC W25 hackathon finals out of 100+ teams.
Voice-first health companion. Voice intake → clinician-style note pipeline (ASR + summarisation) with latency under 1 minute.
Freelance CV project to count small parts in 1:72 HVAC assembly kits from video and image data. Delivered an iOS app for real-time part counting.
Real-time dashboard for whale movements, volume anomalies, and notable trader activity using Nansen on-chain data. Built at Sentient Labs.