About

What I do now

I lead agentic AI systems at a private equity firm. Private equity runs on judgement applied to messy, unstructured information — company filings, management presentations, market reports, hundreds of documents per deal — and most of that work is still done by hand because it demands context and care rather than pattern matching.

The interesting question is not whether a model can read a document. It is where an automated system can be trusted, what it should hand back to a human, and how you can tell when it is wrong. That is a systems problem as much as a modelling one: retrieval, evaluation, guardrails, observability, and the discipline to keep a person in the loop wherever the cost of being wrong is high.

I am also learning the domain itself — deal flow, modelling, how decisions actually get made. Writing about that as I go is part of the blog.

Where I came from

I studied Computational Science and Engineering at ETH Zürich — an MSc built around the idea that simulation, numerics and software engineering are one discipline rather than three. For five years I lived in that world: teaching GPU programming at CSCS summer schools, running the HPC Lab and the C++ course at ETH, and spending a semester at Harvard's CSE Lab.

My thesis built a deep-learning framework for Korali and benchmarked it against cuDNN, oneDNN, PyTorch and TensorFlow. Before that I put approximate Bayesian inference onto GPUs with R-INLA. Both are written up on the projects page.

Then I took it to industry. A contract doing backend and generative AI work, then nearly two years as a Senior Machine Learning Engineer in a Center for Advanced AI — production ML, the unglamorous parts included — and now private equity.

How I think about the work

• Measure before you believe. An HPC background leaves you permanently suspicious of numbers you did not produce yourself. Most "the model is slow" problems turn out to be placement, I/O or memory, not the model.
• The boring parts decide the outcome. Retrieval quality, evaluation harnesses and deployment reliability determine whether an AI system is useful far more often than the choice of model does.
• Automate the work, keep the judgement. The useful target is rarely full autonomy. It is removing the mechanical parts so a person spends their attention where it is actually worth something.
• Build it to find out. This site runs on infrastructure I provision, deploy and operate myself, partly because that is the fastest way to actually understand a thing.

Outside that

Open source, Arch Linux and Sway, Emacs and Vim. Triathlon, climbing, chess. Austrian, based near Vienna, working in German and English.