SiteProof

Rental property inspections produce condition reports that decide who pays for what at move-out, a high-friction, dispute-prone process where a wrong number has real legal and financial weight. The tempting build is "AI that looks at a photo and tells you the repair cost." A property manager or owner would love that. The trap is that an LLM asked for a price will confidently produce one whether or not it has any basis, and a hallucinated repair estimate in a tenant dispute is the kind of error that ends up in small-claims court.

Vision detects, the engine prices. Our core design separates what an LLM is good at from what it must never do. The vision model returns categorical facts only: damage type, severity, material. It's never asked for a dollar figure. A separate deterministic cost engine does the pricing across 62 line items, square-footage based with regional multipliers, validated against 55+ industry pricing sources. The model never touches the pricing path, so it can't hallucinate a number; no prompt or disclaimer required.

The wording is part of the product. Positioning rules are enforced in code, UX, and terms: "AI-assisted, human-verified," never "AI-powered"; "estimate," never "appraisal." Those words carry legal weight, so they live in a three-layer liability framework instead of a style guide.

An inspector uploads photos. The vision layer classifies each detected issue into structured, categorical fields. Those fields key into the deterministic cost engine, which computes a priced line-item estimate with square-footage and regional adjustments. Every figure a user acts on is labeled by how it was produced, so a number is never mistaken for something it isn't. The result is a defensible condition report instead of an opaque AI guess.

I'm building SiteProof with my brother, who leads sales. We're live in pilot today, collecting real uploads and usage to validate the workflow before we scale the cost model to new regions and item types.