About
I’m fascinated by the immense potential of AI — and convinced that unlocking its full potential requires knowing when we can rely on it.
This conviction drives my research. I started by asking a narrow question: can we prove that a model’s predictions won’t change under small input perturbations? This led to certified robustness methods, including the first certification framework for point cloud models (ICCV 2021) and hierarchical certification for segmentation (ICML 2024).
But robustness at test time isn’t enough if the training process itself can be compromised. I extended certification to the training pipeline — proving that models remain robust even when training data is adversarially perturbed (GCPR 2024, NeurIPS 2025). This closes the gap between training and deployment: guarantees that hold end-to-end.
Most recently, I’ve been zooming out. Individual proofs about individual models are necessary but not sufficient. With Scalable Delphi, I’m developing methods to assess the risks of entire AI systems — structured evaluation that scales from expert panels to automated LLM-based assessment, reducing what used to take months to minutes.
Education
M.Sc. in Computer Science — ETH Zürich. Thesis on robustness certification for point cloud models with Martin Vechev at the Secure, Reliable, and Intelligent Systems Lab.
B.Sc. in Computer Science — Karlsruhe Institute of Technology (KIT). Thesis at MPI for Informatics supervised by Bernt Schiele.