Tim-Henrik Buelles

About

I'm a Research Scientist at Pythia Labs. I was trained as a Mathematician, now I train neural nets. More technically (less funny), I currently work on multi-task learning for generative and probabilistic modeling. Previously, I was a Postdoctoral Scholar in Mathematics at Caltech working with Prof. Graber on geometric deep learning and applications in string theory. I did my PhD in Mathematics at ETH Zürich with Prof. Pandharipande. My thesis focuses on generating functions of enumerative invariants derived from Calabi–Yau geometries in superstring theory. Before that, I did my MSc and BSc in Mathematics at the University of Bonn where I worked with Prof. Huybrechts on algebraic cycles and motives of K3 surfaces.

Research

My research was published in Journal of Algebraic Geometry, Forum of Mathematics, and Manuscripta Mathematica.

After spending a few years in pure math, I shifted my attention to the statistical foundations of machine learning. At Pythia Labs, I am currently working on generative models for protein design and drug discovery. We train models on biomolecular data such as sequential 1D amino acid data (autoregressive models, Transformers, Markov chains, etc.), and geometric 3D structure data (Diffusion, Langevin, Graph Transformers, etc.). For each modality or task, we have a well established toolbox for probabilistic modeling, architecture, feature learning, scaling, deployment, etc. The hard part: How to integrate into a single model that can perform multiple conditional generative tasks? Oh, and please make it compute efficient, scalable, robust, fine-tunable, and ready for production..

Teaching

From 2016 to 2023, I was an instructor for Mathematics at Caltech, ETH Zürich, and University of Bonn, teaching a number of seminars and graduate courses.