Field notes, starting

I’m starting this from the middle of the work, not the end of it. Most writing about production AI either reads like a vendor pitch or a victory lap. This is neither. These are field notes — what I tried, what shipped, what fell over in week three.

I’ve spent the last couple of years building agentic systems that have to behave: RAG pipelines that don’t hallucinate citations, memory layers that survive past the context window, multi-agent orchestrations that don’t deadlock the second a tool call times out. Most of them broke the first time. Some of them broke a fourth time after I thought I’d fixed them. The pattern that actually works in production is rarely the pattern in the model card.

What you’ll find here

Three kinds of posts. First, walkthroughs of specific problems — the kind where I show the code, the wrong solution I tried first, and the trade-off I settled on. Second, opinion pieces on tooling and architecture choices that I’d defend in front of a skeptical CTO. Third, occasional updates on what I’m shipping inside Atlas and Nexus.

What you won’t find: think-pieces about whether AGI is near, takedowns of other people’s products, or anything that reads like it was written to game a feed. I’d rather be useful to ten engineers than viral to ten thousand.

A note on tone

Direct, opinionated, honest about what didn’t work. If a technique I’m recommending failed for me twice before it succeeded, I’ll tell you. If I’m uncertain, I’ll say so. The only thing I have to sell here is judgement, and judgement gets cheap the moment it stops being honest.

If any of that sounds useful, subscribe — new field notes go out roughly twice a month. And if you want to put a problem in front of me directly, the work page explains how.