Selected work / 01–02

Questions worth
building around.

Technical projects in robotics and AI systems. The approach, the evidence, and what didn’t work.

Conceptual system illustration

Robotics · Planning under uncertainty

Planning when the world is only partly visible.

Combining offline dynamic programming with online Bayesian belief updates to study how a robot should navigate when it cannot confidently localize itself.

POMDPsBayesian filteringDynamic programming

1.44 → 0.29Bottleneck MAP localization error

PROJECT 02

Conceptual system illustration

LLM systems · Adversarial evaluation

A second model. A wider view of safety.

A conversation-aware safety supervisor that evaluates dialogue history before a chatbot responds, tested across model sizes and moderation approaches.

LLM safetyMulti-turn contextEvaluation

99.5%8B supervisor recall · with over-triggering

Inside each case study

Problem → system design → experiments → results → lessons learned