AI agents for internal operations: practical, unglamorous use cases
The interesting demos book travel. The deployments that survive contract review match invoices to purchase orders. Six jobs boring enough to work, and how to build them safely....
From pilot to production: why most AI projects stall
The team builds something impressive in three weeks, demonstrates it to an approving room, then spends nine months not shipping it. The blockers are always downstream of the model....
Data readiness: what your data needs before any ML project
Every ML proposal says we have a lot of data. It is almost always true and almost never the relevant question. Five checks, and the reason retrieval projects need no training data at all....
How to test software that gives a different answer every time
The same input produces a different answer, so the assertion fails, so someone deletes the assertion. Four scorers in order of cost, and how to fail the build on a real regression rather than on normal variance....
Building an AI feature on your own data without leaking it
Neither major US provider trains on API data by default. That is the settled question. The exceptions in the retention terms, and the leak in your own retrieval layer, are the ones worth your attention....
What it costs to run an LLM feature in production
The model call is rarely the whole invoice. A worked example at 1,000 requests a day, the two levers that actually move the number, and the lines that have no published price at all....
RAG, fine-tuning or prompting: choosing an approach for your data
Knowledge problems are retrieval problems and behaviour problems are tuning problems. How to tell which one you have, and what each approach really costs to run....
Where AI actually pays off in business software
MIT found 95% of enterprise AI pilots showed no P&L impact, and externally partnered builds succeeded at twice the rate of internal ones. What separates the cases that pay from the ones that do not....