In-hand tactile perception
A MuJoCo twin of the LEAP hand with 368-taxel XELA skins. Self-supervised MAE and JEPA models infer in-hand object shape and quantity at over 95% accuracy.
Research from my PhD and visiting positions, a product people use, and the systems I build on the side. Filter by type or by track.
A MuJoCo twin of the LEAP hand with 368-taxel XELA skins. Self-supervised MAE and JEPA models infer in-hand object shape and quantity at over 95% accuracy.
A pipeline that turns egocentric human hand poses into LEAP-hand actions for in-hand multi-object manipulation.
Verified VLA actions for about the cost of one call: 1.1× a plain call versus 20× for CoVer, matching best-of-5 verification with 5.5× fewer policy evaluations.
Can analog thermodynamic hardware replace digital solvers inside a robot’s control loop? In closed-loop MuJoCo pick-and-place, warm-started iterative refinement cut device time about 330× while matching an exact solver.
GRPO fine-tuning of Qwen2.5-Coder with unit-test rewards and a sandboxed executor. Lifted 1.5B pass@1 by +3.2 on MBPP and +5.5 on HumanEval+ (3 seeds), with no reward hacking across 21 audited runs.
Two PyTorch-trained ViTs ported to dependency-free TypeScript with hand-derived backward passes, verified against PyTorch on every tensor.
A transformer built from the ground up in PyTorch, from tokenisation through multi-head attention, to understand exactly how current LLMs work.
A live time-tracking app anyone can use: categories, daily tasks and habits, with a built-in AI assistant and an MCP connector for Claude and other assistants.
A multi-tool agent on Snowflake Cortex that answers sales questions from structured metrics and call transcripts, iterated with observability traces and exposed to Cursor over MCP.
Dense, BM25 and hybrid retrieval benchmarked on a hand-labelled question set. Hybrid lifted chunk-level Hit@5 from 0.66 to 0.78.
Particle swarm and evolutionary search over rack layouts to cut robot collisions in multi-agent path finding.
A reinforcement-learning agent that slices large orders, benchmarked against the Almgren–Chriss schedule.
Orbital mechanics, thermal load, radiation and GPU device physics in one Python simulator.
No projects match both filters yet. Try another combination.