I am a postdoctoral researcher at Imperial College London, working on reliable and efficient ML. My current research examines the stability, safety, and alignment of LLMs and VLMs under resource-efficiency pressures, and how representation geometry mediates this relationship. At Imperial, I have also worked on mechanistic interpretability, learning dynamics, and robustness to spurious correlations. My research has been featured in NeurIPS, ICML, ICLR, ICCV, and Nature Medicine, among others.

Before Imperial, I completed my PhD at Boğaziçi University, where my research focused on robust generalization in deep learning and probabilistic latent variable models. During my PhD, I was a research scientist intern at Google DeepMind, where I worked on human-AI complementarity in medical diagnosis and on evaluating medical AI systems under uncertain ground truth.

I have industry experience in the research, deployment, and team leadership of production ML systems, most notably leading a team behind daily cash-demand forecasting for 8,000+ ATMs.

Updates

• Sep 2026: I will be serving as an Area Chair for ICLR 2027.
• May 2026: At ICLR 2026, we presented our papers on the interaction of compressibility and adversarial robustness and how data correlations shape feature geometry under superposition.
• Apr 2026: Our papers on adversarial vulnerability from interference between features in superposition and multi-way representation alignment have been accepted to ICML 2026.
• Oct 2025: We presented our paper on large learning rates, robustness to spurious correlations, and compressibility at ICCV 2025.