AI Construction Estimator

Construction estimating is slow, manual, and inconsistent. Plan review and quantity takeoff eat hours of skilled time, and the knowledge lives in a few experienced people's heads. The same drawings get measured by hand, over and over, and small errors compound into expensive ones.

Plan review and takeoff are the most repetitive part of estimating: exactly what software should do a first pass on, and exactly what you can't let it do alone. So the tool drafts and the estimator decides. That split is the whole design.

It's built around a real construction business's workflow; the aim is to remove hours of manual takeoff and rework. Any efficiency figure is a target at this stage, not a realized result.

Document-understanding and computer-vision workflows ingest plans and PDFs, run plan analysis, and extract measurements, then assemble a reviewable takeoff the estimator prices against cost data. The goal is a strong first draft an estimator can check quickly, with every figure traceable back to the drawing it came from.

The same idea sits behind my CRM work: a model is useful for drafting and dangerous deciding on its own. The estimator reviews every measurement and every line of the takeoff and can edit or reject anything. A confident wrong number never passes unchecked. The software removes the busywork; the person keeps the judgment. The model is one step in a pipeline built to surface its own uncertainty and catch it when it's wrong.

An early build is mostly open questions. The ones I'm sitting with:

• Where the confidence threshold sits between auto-filling a measurement and flagging it for review.
• How to show uncertainty so an estimator can scan it, not audit every figure. A number and its doubt should travel together.
• Which drawing types to support first. Breadth is tempting; depth on the common cases is probably worth more.
• Whether the honest win is full automation or just a much faster first draft. I suspect the second, and I'd rather build the true thing than the one that demos.

Python · FastAPI · Computer vision · Document understanding · PostgreSQL · React

• An early build in active development, not a finished product; no realized savings claimed.
• Efficiency gains are projected, based on the manual takeoff and rework the workflow is meant to remove.
• Accuracy depends on plan quality and type; ambiguous or low-quality drawings lean harder on estimator review.
• Agentic CRM is the flagship; this applies the same human-in-the-loop philosophy to a different domain.

• Broaden the plan and drawing types the analysis handles.
• Improve measurement-extraction accuracy and confidence signals.
• Tighten the review surface so checking a takeoff is fast.
• Integrate cost data more deeply for grounded pricing.
• Pilot with a real construction business and measure the time actually saved.