I am primarily interested in (multi-agent) sequential decision-making and reinforcement learning. I want to understand how systems can learn from experience, plan ahead, and reason about the consequences of actions in realistic settings with partial observability, uncertainty, and risk constraints.
My MSc thesis studied active feature acquisition as sequential decision-making under costly and missing information. We developed AFABench as a modular evaluation framework and extended the AFA POMDP to missing data during training, showing how missing data can distort both beliefs and the acquisition routes a policy learns to value.
AFABench is a generic framework for benchmarking active feature acquisition in classification. It standardizes episode interfaces, budget protocols, datasets, methods, and diagnostics so myopic, subset-based, and reinforcement-learning policies can be compared under the same cost-performance trade-off.
At AIMLeNS, I worked on scaling generative models for molecular dynamics. This project shaped how I think about generative modeling in scientific settings, where useful models must respect geometry, sampling behavior, robustness, and the physical structure of the system being studied.
Causality-Δ studies how local derivative information can expose dependency structure in flow-matching generative models. We use Jacobian-vector products to trace how latent perturbations propagate through learned dynamics, giving a practical way to inspect model geometry while keeping a clear boundary between local dependence and causal intervention.
My BSc thesis studied uncertainty quantification in computer vision models using Shannon entropy as a practical measure of predictive uncertainty. We focused on how confidence changes under perturbations and when a model output should be treated with caution rather than accepted as a reliable prediction.
