Active-learning loops in construction estimating

Every contractor we've talked to has at least one trade where their drawings use proprietary symbol conventions a generic vision model has never seen. Data-center structured cabling. Industrial process piping. Specialty fire-protection on tilt-wall warehouses. Generic AI gets these wrong, and there is no public training set that fixes it.

The active-learning loop

Our answer is structural: every project produces ground truth. When a vision claim hits the review queue, the estimator either confirms or rejects it. Confirmed claims become annotations. Annotations train a per-org recognizer. Within a few weeks of active learning, the per-org recognizer typically converges meaningfully above the generic baseline.

• Vision claim → review queue (confidence < threshold)
• Estimator confirms or rejects with one click
• Confirmed claim → ground-truth annotation in the symbol library
• Re-train the recognizer overnight with the new annotations
• Next bid sees fewer queue entries — the model learned

The compounding effect matters. Each bid produces hundreds of training examples. Six months in, the recognizer is better at your symbol conventions than any generic model. Twelve months in, you have a moat.