skip to content
Serafim Sukhenkiy

What do I do? Stare at problems until I understand them, then build something. Sometimes it takes a week. Sometimes three years. I'm not great at estimating. Terrible at planning, actually. But good at not giving up.

I think about infinity more than is probably useful. Questions that don't have answers. That's also a kind of poem - the ones that say something true in a way that shouldn't work but does.

Got too many opinions. Many aren't popular. Not always right, but always interested.

But talk is cheap.

Posts

I rewrote my prompt five times. It got worse every time. - The system prompt was the least managed file in my project. Here's what finally made me measure it.

My staging is a full copy of prod. It costs nothing extra. - Staging is usually bad because it costs the same as production and earns nothing. Owning the hardware removes that argument.

Why I run Kubernetes without EKS - Renting compute is expensive. Cloud reliability is worth paying for. You can have both, in one cluster.

A good LLM harness hides the failures it handles - My error logs were clean for two weeks. My agents were failing the whole time. Both were true.

Code is cheap. Decisions aren't. - Code used to be the expensive part. It isn't anymore.

OCR in 2026 - still a problem? - OCR is supposed to be solved. Then I tried to read engineering drawings.