Research
My research is organized around one question: how can learning systems make decisions under uncertainty in a way that is risk-aware and verifiable? The four directions below develop complementary pieces of this question — risk-aware learning provides the objective, verification and reliability provide the guarantees, LLM decision systems test the methods on language-driven control, and power-system applications ground the work in a reliability-critical domain. Papers are intentionally cross-tagged across directions, since most of them advance more than one.
Themes
Each direction includes multiple related papers and paper-derived visuals.
Methods for learning decisions under uncertainty, distribution shift, and explicit risk specifications.
This thread asks how policies should adapt when the cost of failure changes across environments, users, or operating regimes.
Stress testing, verification, and input-space construction for learned models used in constrained settings.
This direction focuses on making model behavior inspectable before learned components enter reliability-critical workflows.
Fine-tuning and inference-time methods for language models that propose technical actions under constraints.
The goal is not only better text generation, but more reliable decisions when language models interact with structured domains.
Power-system testbeds for studying learning, optimization, reliability, market participation, and control.
These projects use power systems as a concrete domain where uncertainty, constraints, and model reliability matter immediately.