Calibration and decisions
How should a system act when context changes, evidence is incomplete, or confidence is unreliable? My earlier T-UEBA work informs my interest in calibration, abstention, and analyst review.
Agent evaluation
Designing tasks and feedback that reveal why an agent failed: a bad tool call, a missing observation, or a flaw in the scoring rule.
Geometry and representation
Combining exact geometry with learned relations to reconstruct CAD while preserving editability.
Robot planning
Comparing motion plans against geometric constraints, smoothness, reachability, and operator expectations, then testing their behavior in the complete system.