A working definition, mostly for myself. Expect this to change.
When I say AI factory, I don't mean a model, and I don't mean a chatbot bolted onto an IDE. I mean the production system around AI: the repeatable pipeline that turns models, data, context, and evaluation into reliable developer value — over and over, safely, at the scale of a whole engineering org.
A factory is the right metaphor because the interesting problems are industrial, not magical:
- Throughput — how much useful work moves through, not how impressive a single demo is.
- Quality control — evaluation and guardrails as first-class infrastructure, not an afterthought.
- Standardized parts — shared context, retrieval, and tooling so every team isn't rebuilding the same jig.
- Feedback loops — measuring whether any of this actually made developers faster, and feeding that back in.
The questions I keep circling:
- What's the unit of output? A merged PR? A resolved ticket? Saved developer-hours? You optimize what you measure, so this matters.
- Where does human judgment stay in the loop, permanently — not as a temporary crutch?
- How do you keep the platform opinionated enough to be useful but open enough that teams don't route around it?
Related: Start HereStart Here
Welcome to my digital garden. Unlike the blog, nothing here is finished — these are notes I tend and revise over time. Some are a paragraph; some will grow into essays; a few will quietly die. That.... This connects to a lot of what I write in the blog.