When we escalate
Confidence under 75%.
Roughly 7% of submissions. We route to a human inspector. You get an annotated response in 24 hours.
We trained a vision model on tens of thousands of post-build defect photographs — labeled by structural engineers, warranty adjusters, and builders themselves. It tells you what you're looking at, how confident it is, and what to do next. You're still the human. We just remove the guesswork.
We did not train this model on scraped images. We built our training corpus from working warranty case files — every photograph reviewed by a structural engineer and a builder, with the final disposition (covered / not covered / monitor) attached. The model learns what humans actually decide, not what looks dramatic.
For every photograph, the model produces a verdict and its strongest counter-argument. If it predicts "settlement crack — claim-worthy" at 92%, it also surfaces "this could plausibly be drywall shrinkage in a corner" with the evidence and how to disambiguate by re-photographing.
That second voice is what an inspector does in their head. We just made it visible.
Any model that pretends to never be wrong is a model you can't trust. Ours flags any image where confidence dips below 75% as "escalate to a human reviewer." Those go to our review queue and come back with an annotated explanation within 24 hours — at no extra cost during your first year.
Roughly 7% of submissions. We route to a human inspector. You get an annotated response in 24 hours.
Diagonal cracks > 1 cm, foundation displacement, beam deflection. Always referred to a licensed engineer in your state.
We draft language and assemble evidence. The conversation between you and the builder is yours.