I build AI systems, then figure out where they break.
The demo is the easy part. I care about what happens next.
About
I’m a software engineer working across backend systems, retrieval, agent workflows, data pipelines, and evaluation: the parts that turn an AI idea into something people can actually use.
Right now I’m building multilingual LLM workflows at Infineon, automated peer-review agents at ETH Zürich’s Agentic Systems Lab, and research on model behaviour at RC2S2, ELTE.
I keep coming back to failure. What does an agent do when a tool breaks? How do you notice retrieval returned the wrong document? How do you keep a complicated system understandable?
I build the thing, watch it fail, then make it less fragile.