An estimating engine learned the difference between confident and correct

The problem

A remediation contractor prices one to two thousand estimates a year from messy field notes and photos. The first-generation quoting tool matched the benchmark job perfectly, then live validation told the truth: on real-world notes it extracted the wrong quantities and confidently mispriced most jobs, sometimes by thousands of dollars in either direction. Fast and wrong is worse than slow.

What we built

• An evidence-backed extraction engine: the AI reads the job notes and photos and must cite what it found before anything gets priced. No grounded quantity, no price.
• A held-out validation set built from the contractor's own history: real past jobs with real final prices, so every change to the engine is measured against reality instead of vibes.
• Seven scored evaluation rounds teaching the engine the estimator's pricing conventions, with a report card after every round.
• An automatic-draft gate: the engine only writes an estimate on its own when extraction is fully grounded. Ambiguous jobs route to a human with the evidence attached. In validation, the rebuilt engine prevented every one of the fourteen bad auto-adoptions the old version would have made.
• A correction-capture flywheel: every human fix to a draft is logged and becomes evaluation signal, so each round of tuning is grounded in what the estimator actually corrected.

Stack