Prompt and Pray

Agentic CEO

Sat, 16 May 2026 00:00:00 +0000

Sitting at my kitchen table, on an uncomfortable chair, in the middle of the day, my laptop in front of me, for no good reason whatsoever, I have decided turn AI into my manager CEO.

You may be wondering about the motivation. It is quite simple. I have a ton of unfinished projects. It’s that simple. Over the years, I have started a lot of software projects. Some of them are half finished, some are done. Some were personal utilities, some could have been full successful startups. What they all share in common is that none were shared with the broader audience. The reason was simple as well. I was always constrained for time. All of these projects were side-projects. They were developed over prolonged time-horizons, after work hours and on weekends. The agentic experience changes the calculus of time for me. It makes a lot of things possible again without having a dedicated team, fund raising or hiring people to help out.

About

Sat, 09 May 2026 00:00:00 +0000

This blog is a field journal, not a polished publication. Expect rough edges, failure reports, whacky ideas and the occasional thing that actually worked.

What I write about:

Eval experiments — testing model claims against real tasks
Tooling and workflows — building with LLMs, not just talking to them
Lab notes — shorter observations that don’t warrant a full post
Math - occasionally, I dabble in mathematics and write some stuff here

If something here is useful to you, I’d love to hear about it.

Model self-checking

Sat, 09 May 2026 00:00:00 +0000

Giving coding agents an ability to self-check their own work is one of the most useful techniques I came across. I develop a lot of numerical software in my free time and most models are very helpful here but no model has earned my complete trust. They simply don’t consult with me presence of uncertainty and rely too much on the parametric knowledge instead of consulting external sources. The usual runaround happens when the model states that it knows the answer to some problem, codes the algorithm, and the algorithm looks plausibly correct but doesn’t really work. A lot of debugging follows and the result is a lot of wasted time and frustration.