Projects
Machine Learning & Decision Systems
-
Offline RL that learns when not to act. FQI distilled into a depth-10 tree with guardrails; value per thread from −13.46 to +15.25.
Computer Vision & Perception
-
Where should the suction cup land? FastSAM proposals, a 2-of-3 tote/item vote, and a distance-transform pick point: 0.958 pick score vs 0.31 baseline.
-
Learned features survive motion blur where ORB and SIFT do not; a factor graph lets loop closures correct drift retroactively.
Privacy & Trustworthy ML
-
Three DP releases budget-capped at ε = 3, a differencing attack that shows why the cap matters, and an ablation that proved my own budget hypothesis backwards.
Agentic AI & LLM Systems
-
A skill does not merge unless an eval proves it helps. 17-skill registry, 466-task bank, 3-tier CI gates.
-
Agents hire and pay each other on-chain. Reputation is bonded capital, so cheating is economically punished rather than merely recorded.
-
Raw feeds to readable digest. Zero-shot classification means the topic taxonomy can change without retraining.
Developer Tools & Open Source
-
LLM agents control a game engine. The tool boundary is the security boundary, so it has to hold even when the model is wrong.
-
User endpoints were silently overridden by defaults. Fixed resolution order, then added quarantine so one bad provider cannot degrade everything.