Moake Park Group × Fort AI Agency
AI Construction Document QC
Cross-references your specification manual against the drawing set and surfaces every mismatch — equipment specified but not shown, fixtures shown but not specified, finish callouts that contradict the schedule, fixture counts that don’t match submittal quantities. Each finding is page-cited on both sides so a reviewer can audit it in under a minute.
What the engine flags
Equipment specified, not shown
Spec calls out a piece of equipment, drawings don't show it on any sheet. The most common QAQC catch — and the most expensive to miss after bid.
Drawn but not specified
Fixture appears on a plan or schedule but the spec doesn't define it. Submittals will get rejected; field installs the wrong product.
Quantity mismatch
Spec says 12 toilets, drawings show 14. Bid quantities and submittal counts get caught before they hit the contractor.
Finish callout conflicts
Room finish schedule says VCT, plan tag shows carpet. Coordinate before bidding to avoid change orders.
How it works
- 1
Ingest
Upload the spec manual and all discipline CD sets. The system splits, OCR-checks, and indexes by CSI division + section.
- 2
Extract
Per CSI section, the engine builds a structured catalog of every spec'd item — equipment, fixtures, materials, finishes — with verbatim evidence quotes and absolute page references. Drawings get parsed for schedule tables, plan callouts, and sheet indexes.
- 3
Reconcile
Spec items and drawing items are cross-referenced using model numbers, tag marks (P-1, AHU-1, WC-1), descriptions, and room locations. Every mismatch is flagged with severity, confidence, and two-sided evidence.
- 4
Review
Output is a structured QC report — each finding has the spec page + sheet number, verbatim quotes, and a recommended next action. A reviewer can audit each flag in under 60 seconds.
Accuracy is the whole point
QC tooling only earns trust if every finding is auditable. The architecture is built so each flag carries everything a reviewer needs to verify or dismiss it in one screen.
- Every spec item is anchored to a verbatim evidence quote and an absolute page number in the spec manual.
- Every drawing item is tied to a sheet number and a source (schedule table, plan callout, legend, detail).
- Every finding includes the spec quote, the drawing reference, and a confidence rating — never just an unsupported assertion.
- When evidence is one-sided, the engine outputs status: indeterminate — never a false flag.
- All inference runs through enterprise Anthropic APIs under terms that prohibit training on Moake's documents.