Field proof
The Drawing Review, Automated
WO-004 · AI extraction · Precast Concrete- Situation
- A precast concrete supplier whose work starts where a civil drawing set ends. To quote a job, their estimators need to know every drainage structure in it — every catch basin, manhole, junction box, headwall, and pipe segment.
- What they tried
- Hundreds of pages of civil drawings, one structure missed on page 147, one bad bid. Now the AI finds them, proves them on the drawing, and measures its own accuracy. A drawing set arrives — often hundreds of pages of site plans, drainage profiles, and schedules. An estimator opens page one and starts looking. Structure by structure, page by page, cross-referencing the plan view against the profile against the schedule table, writing down IDs and elevations by hand. The same structure appears on three different sheets under three slightly different labels, and reconciling them is on the human. It's slow. It's brutally repetitive. And it depends entirely on attention — a structure missed on page 147 of a 300-page set is revenue that never gets quoted, or a bid that's wrong. Twenty years of software shopping never found anything for it, because no vendor was ever going to build a product for this exact problem. It's the market of one, in its purest form.
- What we built
- An AI system that reads drawing sets the way an experienced estimator does — and then proves everything it found. Reading the whole set: upload the PDF, pick the pages that matter, and the pipeline runs in the background — it classifies each sheet, discovers the structures, pulls the rim and invert elevations and pipe data, and reconciles what it found across views. The output isn't a summary — it's a structured inventory of every structure and pipe segment, exportable to Excel, PDF, or straight data. Reading the way the drawing was meant to be read: drawing sets carry their own cross-checks — the same structure shows up in the plan, the profile, and the structure schedule. The system uses that the way a careful human would — it reconciles its findings against the drawing's own schedule table, recovering anything it missed, and treats CB-1, CB01, and CB 1 as the same structure the way any estimator would. When multiple views confirm the same structure, its confidence goes up. Corroboration is the method, not an afterthought. An AI that doesn't trust itself, on purpose: every structure the AI claims to have found is checked against the document's actual text; a claim the document can't corroborate gets its confidence cut before a human ever sees it. Known noise — grade callouts that look like structure IDs, placeholders, unreadable phantoms — is filtered by rules tuned from real failure cases. The system's default posture toward its own output is suspicion, which is exactly the posture you want from software doing estimating. Showing its work: every extracted structure carries a confidence score and links to its source page — highlighted with a box on the actual drawing, so verification is a glance, not a search. The estimator's job inverts: instead of hunting for structures, they confirm them. Measured, not promised: the system includes its own accuracy instrument — feed it the ground truth for any job and it reports precisely what it caught, what it missed, and what it over-found. We don't estimate the system's accuracy. We measure it, per job, and tune against real failures — which is how the second generation of the extraction engine came to read the document's own text first and call on AI only where the text runs out. Faster, cheaper, and more accurate than the first, because the tool is allowed to evolve the way the work demands. And a feedback loop the customer owns: when something's missed, flag it — an assistant helps refine the extraction instructions, so the system keeps getting better on their drawings, not a vendor's test set.
- How fast
- Verification now runs in minutes per drawing set instead of hours of page-by-page review.
- What changed
- A review that consumed hours of an estimator's attention now takes minutes of verification. Structures stopped slipping through on page 147. And the company owns a tool that no one was ever going to sell them — built for the exact problem their business runs on.
- What's next
- With manual review eliminated, the team is looking at where else this kind of automated drawing extraction fits across their estimating process.
What's the workflow nothing fits?
The form your team fights with. The process living in a spreadsheet. The tool you pay for and route around. Tell us about it — that's where we start.
Tell us about your workflowLive in 30 days or it's free. Cancel anytime. Keep everything you own.