My research focuses on reliable and data-efficient reinforcement learning for sequential decision-making under uncertainty. I study how imperfect but useful structure, including coarse predictions, decomposable dynamics, optimization geometry, stability, and operational constraints, can guide learning when data are limited and failures are costly. A central goal is to develop algorithms with rigorous guarantees on sample complexity, performance, and constraint satisfaction. I apply these ideas to energy systems and AI computing infrastructure, where online scheduling, control, and policy design coordinate physical resources and computing demand under uncertainty.