I started this work because a diagnosis made the timeline personal.

Someone close to me was diagnosed with type 1 diabetes. That is the reason I moved from machine learning as a research interest to machine learning as a way to make discovery go faster. If a system can read the literature, form a hypothesis, test it against something that can actually check the answer, and show its work — then the rate at which we find things is no longer bounded by how many people are reading papers.

The obstacle is that today's models are fluent and unreliable in the same breath. I have spent the last few years on the same idea from several directions: don't ask a language model to be right, ask it to produce something you can check. Use it as a semantic parser that turns messy intent into typed structure, then hand that structure to a solver, a type system, or a test — something with no opinion and no charm.

What that looked like in practice

At JKU Linz under Sepp Hochreiter it was the parameter-choice problem: how do you pick a hyper-parameter in unsupervised domain adaptation when you have no target labels to validate against? We proved you can aggregate models and land within a factor of two of the unknown optimum. That got an ICLR Oral, but the useful part is the shape of the answer — a guarantee instead of a heuristic.

At ExtensityAI it became SymbolicAI: an open-source framework where generative calls are mediated by contracts, with pre- and post-conditions bound to schemas and bounded automatic repair before anything escalates. That framework is the reason the company could sell multi-step generation to people who could not tolerate a confident wrong answer. It also got a fair amount of attention, including from people who had spent years arguing that pure scaling would not be enough.

At Alpha Omega Labs it is infrastructure. Once agents write and deploy real code, the interesting questions stop being about model quality and start being about governance: who is allowed to run what, against which systems, with what audit trail, and what stops a run from quietly spending money all night. Cortex is that control plane. omegaXiv is what you build on top of it once you trust it.

How I work

I still read and write the code. I have found that AI leadership decays quickly into slideware when the person doing it stops being able to review a pull request, and the decisions that matter — which abstraction, which guarantee, which thing to not build — are not visible from the top of a roadmap.

I am biased toward enforcement over convention. If an architectural rule matters, it should fail the build, not live in a document nobody reads. Both platforms I run check their own layering in CI, and their test harness treats a skipped test as a failure, because a green suite that quietly stopped asserting anything is worse than a red one.

Where I work best

My work sits where the research agenda and the production system are the same conversation — the ground a Chief AI Officer, Head of AI Research or principal technical lead covers. In practice that has meant owning a research direction and the platform that ships it, rather than handing one to the other.

The problems I am most useful on are hard enough to need real research and consequential enough that the output has to be right: regulated industries, scientific tooling, anything where a plausible-sounding wrong answer costs something.

Based in Austria, used to distributed teams. Email if any of this is relevant to what you are building.