PhD Candidate, Computer Science

Designed grid-world delivery experiments showing that behavioral implicit communication (BIC) under an inaccurate capability model triggers a "conformity trap": humans defer to the robot's signaled plan, reinforcing its wrong assumptions. Developed a capability-elicitation framework that recasts alignment as an exploration–exploitation problem, where targeted probing accelerates convergence to an accurate human model and improves team performance.

Built the GAMBiT Theory-of-Mind Defender Agent, a PsychSim-based model that infers a network attacker's latent cognitive vulnerabilities from behavior alone. It maintains alternative explicit mental models, each a candidate bias (loss aversion, base-rate neglect, confirmation bias, sunk-cost, availability) with its own utility function, and updates a probability distribution over them as it observes attacker actions, enabling counterfactual reasoning to drive deception and defensive strategy. Validated in human-subject experiments (n = 61) in a simulated enterprise network.

Built a web-based escape room and ran an affect-labeling vs. control study identifying behavioral inflection points where unresolved negative affect predicts behavioral collapse, while resolved affect tracks efficient completion, grounding the design of affect-aware interventions that detect bounded-rationality signatures under emotional load.

Co-designed the Virtual Human Improv Library, a multi-agent (Orchestrator / Director / Actor) LLM framework for social-skills training through improvised role-play, grounded in Stanislavsky's Active Analysis and framing improvisation as real-time Theory of Mind practice.

Analyzed a 10-month LLM-based ticket-routing deployment at a fintech firm to show that automation relocates rather than eliminates articulation work, surfacing how teams negotiate ownership of ambiguous work.