What we ran into, what we changed, and what we would do differently. Mostly written the week it happened.
Latency, refusals, and the moment a candidate realises they are talking to a model.
Tracking how answer engines describe a company, and what actually moves a citation.
Schemas, guardrails, and why generated pages need the same review as human ones.
Confidence, sources, and undo. The three controls every AI interface needs.
Small, boring test sets beat clever prompts every single time.
Token maths, caching, and the point where a demo becomes a bill.