Writing
Blog
Notes on AI product execution, autonomous workflows, and building useful systems that survive contact with reality.
I tested whether Jev’s uncertainty is real
Typed judgments return a distribution instead of a string. Six runs on one deliberately ambiguous ticket to find out whether that uncertainty is real signal or decoration.
Read article
Price it before you build it
AI unit economics as product work: unit costs per route, cost as a ship gate, and why "prices will drop" is not a strategy.
Read article
Evals are your QA team now
Traditional QA cannot test LLM features. Eval gates can: reliability, safety, and regression control as ship gates, and the loop engineering that makes them fast.
Read article
Stop prompting AI. Start defining done.
The useful shift is not better prompting. It is writing loops with goals, checks, retries, and budgets so the machine can verify progress without your attention.
Read article