Selected Work
A selection of my work. Much of my research and engineering at Google DeepMind is internal, so these are the published highlights, alongside my Master's research at ETH Zürich.
Google DeepMind
Senior Research Engineer, in collaboration with dozens of amazing colleagues
- AlphaProof Nature 2025
Core contributor. Created its test-time RL (TTRL), the inference method that solved 3 of 6 IMO 2024 problems — the first (silver) medal-level AI result at the IMO.
- Formal Conjectures Preprint 2026
Co-first author. An open Lean benchmark of 1000+ open and diverse conjectures for verified mathematical discovery.
- AlphaProof Nexus Preprint 2026
Contributor. AI-driven formal proof search for research-level mathematics; resolved nine open Erdős problems.
- IMO 2025 Gold 2025
Core contributor. End-to-end natural-language solutions at gold-medal standard under human competition conditions.
- Gemini 3 Family 2025–2026
Contributor. Post-training across the Gemini 3 family (notably 3.1 Pro).
- SynthID-Image Technical Report 2025
Contributor. Image watermarking deployed at internet scale (10B+ images and video frames).
ETH Zürich
Master's research at the SRI Lab, with my wonderful collaborators Mark Niklas Müller, Marc Fischer, and Martin Vechev
- Boosting Randomized Smoothing with Variance Reduced Classifiers ICLR 2022 (Spotlight)
First author. Certified robustness via variance-reduced deep-learning ensembles.
- (De-)Randomized Smoothing for Decision Stump Ensembles NeurIPS 2022
Co-first author. Certified robustness for tree-based models via (de-)randomized smoothing.
- Robust and Accurate — Compositional Architectures for Randomized Smoothing ICLR 2022 Workshop on Socially Responsible ML
First author. Compositional architecture (ACES) that mixes smoothed and standard models for robustness with high accuracy.