PROJECTS / RESEARCH PROGRAMS
Research programs that connect individual papers, systems, and theoretical questions into longer technical agendas for trustworthy AI.
2026 - Present
An evidence-governed kernel for long-running and self-improving agents.
ProContract separates an agent's adaptive policy from a small trusted kernel that controls obligations, evidence, and canonical state transitions.
2024 - Present
Optimization and search methods that turn formal guarantees into practical verification systems.
A research program on stronger relaxations, tighter domain reduction, and more informed branching for complete neural network verification.
2025 - 2026
Separate targeted forgetting from capability preservation through optimization geometry.
A geometric approach to language model unlearning that controls interference between forgetting objectives and retained behavior.
2023 - 2026
Control and reinforcement learning that hold up when the model was fit from limited data.
Applying distributionally robust optimization to stochastic control and offline reinforcement learning, in the cases where the robust problem stays tractable enough to solve.
2025 - 2026
Simulating how a regulatory alert moves a supply chain when nobody sees the whole picture.
An LLM-agent simulation of drug shortage dynamics under information asymmetry, calibrated against a dataset of 2,925 historical FDA shortage events.