Robotics & ML engineer
I work on the unglamorous middle — the part where a model meets a physical system, or a benchmark, and the assumptions stop holding.
Trained a 29-DOF humanoid to stay upright on alpine terrain, then gated every policy on a second physics engine that shares no code with the trainer.
Lead engineer at Geologic Dome. Reinforcement learning, domain randomization, and the sim2sim validation gate.
Turned monocular video of real trail into an Isaac Sim collider a humanoid can train on, and scored the reconstruction against survey-grade GNSS ground truth.
The reconstruction wrapper, the cleanup chain, the benchmark harness, and the terrain build. Deliberately a starting point: world generation for RL simulation is the area this work narrows toward, and the backbone evaluation continues past LingBot-Map.
Asked how compressible a VLM's visual-token path is, found 50–85% of it removable for under 2% accuracy, and published the hypothesis that failed.
The whole study: instrumentation, five scoring methods, the allocator, and the paper draft.
Fine-tuned a YOLO detector to track endangered great apes — 874 individuals across 500 camera-trap clips — and prepared it for deployment on autonomous field nodes.
Detector fine-tuning, tracking pipeline, and the full-dataset evaluation.
Open to work on physical AI — simulation, locomotion, and the tooling in between.