Risk-aware machine learning and RL
Developing methods that account for uncertainty, distribution shift, and downside risk when policies move across environments or operating regimes.
Research
Risk-aware machine learning for reliable decisions.
My work studies learning methods that remain useful under uncertainty, constraints, and distribution shift, with applications in RL, LLMs, and power systems.
Themes
Current directions
Data-driven reliability and verification
Building learned models whose predictions and decisions can be stress-tested, verified, and trusted under realistic constraints.
LLMs for constrained decision-making
Studying how large language models can support reliable actions in technical domains, including power systems, control, and optimization.
Selected Work
Recent projects
CAT
Caution-aware transfer in reinforcement learning via distributional risk, focused on robust policy transfer under uncertainty.
SITAlign
Bounded rationality for LLMs through satisficing alignment at inference time, accepted to ICML 2025.
LEVIS
Exact and verifiable input-space construction for neural networks, accepted to ICML 2025.
LLM fine-tuning for constrained actions
Fine-tuning large language models to generate economical and reliable corrective actions, with power systems as a testbed.