Vladimir Popelygin

I design AI systems for sports education, adaptive learning, reporting and analytics. We launched the first one in 2024. We do not take the model’s word for anything there: every citation in an answer is checked against its source.

Separately, I work on agents. The infrastructure I am responsible for goes through a PCI DSS audit every year as a service provider, and coding agents there follow the same rules as people do. I am building a safe perimeter for them, and also a data agent.

My sense is that the more work goes to agents, the more it matters to be able to check what exactly they did. That is what I want to spend the next few years on.

What I build with my own hands. How to let coding agents work where all the secrets and customer data live without losing either. Why a data agent that answers questions about data is more dangerous than any chatbot: a plausible wrong number is worse than an honest refusal. When a small model or a plain logistic regression is enough, and when you really need a large model. How to evaluate agents and tell which change breaks what.

A separate thread is building software together with agents: how to give agents tasks, where they break, and what has to be added to the process.

The writing is in progress, mostly in Russian. The easiest way to follow it is rss.

Married, three kids — but this blog isn’t about them.

I photograph the nature around me and city landscapes, mostly on my phone, and when I get the chance I take my Nikon DSLR along. My favourite is an autumn sunrise, when the light breaks through the fog and time seems to stand still. When everything around changes so fast, that feeling is what I miss most.

I read a lot and widely, but what I love most is the classics: Dostoevsky, Tolstoy and, outside Russian literature, Remarque. The last book that really got to me was John Williams’s Stoner. I think it found me at the right time and the right age.