The AI That Held

A home inspection company owner had already tried building this himself. Hundreds of hours into ChatGPT versions, all of them drifting. The rules got lost somewhere in one long chat and the model started improvising. That is how bolt-on AI usually ends, and it is why he came to me instead of trying another chatbot.

I built the v1 engine. It reads his real inspection reports end to end and returns the review in his own scorecard format, the layout he used to fill in by hand at night. His standards are the grounding, so it flags things the way he does, not the way a generic chatbot would.

The model was never his problem. In my build the rules live outside it. Each check hands the AI only the slice of his standards that one question needs, so it cannot lose the plot the way a chat window does. A prompt is not a system. The machinery around the model is what makes AI dependable inside a real business.

I could quote an accuracy number today and it would look great. It would also be close to meaningless. The reports I have are the ones he already corrected, and testing on those flatters the engine. I won't publish an accuracy number until it's measured against reports the way his inspectors first submitted them. When a vendor hands you a suspiciously clean number, ask what they tested it on.

AI implementation · the client's own standards wired in as grounding · per-check retrieval · structured scorecard output