Kanopi · Knowledge Base Common Ground · Reference

A takeoff is the count: how many doors, how much drywall, how many linear feet of footing. Traditional takeoff is hours of manual click-to-measure work. Autonomous takeoff does it from the plan PDF.

The pipeline

Kanopi's takeoff runs as a fixed sequence, from raw PDF to assembled bid, in two to four minutes of wall time:

1 · Ingest and classify
The PDF is rendered at proper resolution, OCR fills text gaps, and every sheet is classified: site plan, floor plan, elevation, roof, schedule.
2 · Extract
Vision models read each sheet for quantities; a schedule parser independently pulls door and window counts from the document text.
3 · Cross-verify
Independent sources must agree: schedule against vision, model against model, pass against pass. Disagreement is surfaced, never silently averaged.
4 · Price
Verified quantities run through the ratio engine and the calibrated rate library for the project's track and tier.
5 · Assemble
Line items build up in CSI MasterFormat divisions; one takeoff can produce a conventional bid and a panelized-system variant side by side.
6 · Output
A branded PDF, structured data, and a push into the project tracker, with contract-type verbiage for fixed-price, cost-plus, or time-and-materials.

Why verification matters

If the counts are wrong, every downstream number is wrong, so takeoff is treated as a stack of verification layers rather than one clever trick. Cheap, high-value checks come first: clean input and sheet classification, then cross-source agreement between the schedule and the floor-plan read, then self-consistency from repeated reads of the same sheet. More expensive checks follow only where needed: consensus between two independent vision models on contested quantities. Construction sanity rules (such as square-footage-to-bedroom ratios and door-count boundaries) catch what statistics miss.

Honest uncertainty

The working target is roughly 95% autonomous with 5% human review. Chasing total autonomy on construction documents is a fool's errand, and a calibrated system is honest about that: when models disagree, the disagreement appears on the operator's screen as a flag to resolve, not a quietly averaged guess. An operator can drill into any line and see its confidence, its rate key, and the formula in plain English; overrides require a reason and persist.

Confidence on every line

Each line in the finished bid carries a confidence tag: quoted (a sub wrote the number for this plan set), takeoff (measured from the drawings), assumed (a market average or ratio), or allowance (a placeholder expected to move). The tag prints on the document, which turns the bid from a lump-sum total into something auditable line by line.

Related