Co-Founder · Cybernetic Labs
Mar 2026 – Present · San Francisco, CA
- Designed and shipped RoboEval's VLM-as-Judge evaluation system (dual-judge mode, Cohen's kappa agreement; 86/86 tests passing).
- Trained and evaluated a fine-tuned SmolVLA vision-language-action policy on the SO-101 arm for pick-and-place and cube-stacking.
- Built PackingBench, a simulation benchmark for embodied / VLA manipulation policies on warehouse packing tasks.
- Conducting mechanistic interpretability research on VLA policies to probe and explain model internals.
- Extended the sim-to-real pipeline (2-camera SmolVLA deployment, joint-space transfer) and the FastAPI/React evaluation dashboard; collaborated with the CEO on the pre-seed deck.