About

I develop learned surrogate models of complex physical systems, alongside interpretable approaches such as symbolic regression and Kolmogorov–Arnold Networks. My work also covers automated workflows for data preparation, training, hyperparameter optimization, and model evaluation.

I worked on two applications of machine learning: web automation and home energy management. I developed a vision-language model for identifying webpage elements and integrated it into a prototype autonomous navigation agent.

For energy management, I implemented and evaluated predict-and-optimize methods in a simulated home environment, studying their performance and failure modes.

My doctoral work explored reinforcement learning and optimization under uncertainty. I developed the Max-Reward framework for goal-reaching tasks, published at ICML 2024, and methods addressing the zero-gradient problem in decision-focused learning with convex optimization.

I also worked on electric-vehicle charging under partial observability, validating the optimization approach on a physical DC microgrid.

I built deep-learning pipelines to detect and classify naked mole-rat vocalizations, together with Python tools for feature extraction and large-scale sound analysis. The automated pipeline reduced manual annotation effort by approximately 80%.

These analyses contributed to evidence that naked mole-rat colonies have distinct vocal dialects, reported in Science.