Jeremi Kaczmarczyk

I've worked on applied machine learning since 2017, mostly on one problem: getting models to do useful work and measuring honestly where they fail. At Tooploox that meant reinforcement learning for refinery scheduling with AVEVA, which became a patent application, and a credit-scoring model that looked like 0.88 AUC until the evaluation respected time order, when it dropped to 0.64. My master's thesis stayed on the same theme: evaluating investment strategies built with machine learning, with a backtesting framework that handles the time structure of financial data.

After that I built backend and data systems: a data platform for a Swiss healthtech company, then tech lead for the rebuild of Affirm's transactional messaging platform, behind billions of messages a year.

Now I build software with coding agents and work on making that reliable: harnesses, evaluation, and deciding which agent-written changes should ship. I founded Veztral to work on this. Silimare, the quant research project I co-founded in 2018, is where I try it first.

I studied architecture before any of this. This site is where I write down what I build, what I read, and what changed my mind.