Topic
What AI genuinely accelerates in delivery, and where it does not.
9 articles
Designing a retrieval assistant to run on the same host as the application, on two shared cores with no accelerator, and to fail closed rather than fall back.
A marker protocol let one reply carry both chat and a draft. Truncation delivered the opening marker without the closing one, and the raw marker reached the screen.
A pinned model returned 404 for every newly created key while still appearing in the provider's own model list. Availability is a per-account fact.
Reviewing 362 business-authored pricing rules needed an ordering, not an oracle. Keeping the AI verdict in its own column is what made it safe to use.
Ask a model for fifteen items and you may get twelve. The choice is between failing, quietly shortening the list, or returning the twelve and saying so.
A fallback chain that advances on error codes but not on empty responses is not a fallback chain. It is one provider with extra configuration.
Embedding existing labelled data into class prototypes gives you a classifier in an afternoon — and a confidence margin that knows when to ask a person.
A model that cannot see a tool cannot call it. Building a natural-language interface where authorisation is structural rather than a prompt instruction.
Where AI genuinely speeds up software delivery, where it quietly costs more than it saves, and why review discipline must not loosen.
