Homebuilding
Drafters spent roughly 20 hours a week reading a community's architectural guidelines and comparing them against a house drawing, by hand. We built an agentic AI product that reviews both documents and surfaces exactly what needs to change.
Users
Drafters & architectural teams
Before
20 hrs/week per drafter
After
a quick review
Every community publishes its own architectural guidelines, covering garage doors, exterior materials, roof details, setbacks, and dozens of other requirements.
Drafters had to find each applicable requirement and manually compare it against a reusable house drawing called a Master Model.
The real bottleneck wasn't editing the drawing. It was discovering what needed to change.
Upload the guidelines and the Master Model. The system extracts the requirements that apply and returns a structured comparison instead of a single verdict, with every finding linked back to the source passage so a drafter can check it in seconds.
Compliant
Already reflected correctly in the drawing.
Non-compliant
A likely change the drafter should review.
Indeterminate
Needs human judgment or more information.
01
Construction documents are ambiguous. An indeterminate state prevented uncertain conclusions from being presented as facts.
02
Source references let drafters validate findings without searching through the original documents.
03
The system accelerates document review; it never edits the drawing or replaces professional judgment.
Information gathering
20 hrs/wk→a quick review
Drafting capacity
4 jobs→6 jobs
Approval rejection rate
60%→40%
The system has to reason across two very different inputs: long, inconsistently structured guideline documents, and house drawings full of dimensions, elevations, and design symbols. It grew from a single-model workflow into a multi-agent system: document chunking, dedicated extraction and validation agents, parameter-specific comparison logic, and design-example matching, all built to keep every finding traceable back to its source.
The product became useful not when it could produce an answer, but when it could show its evidence, admit uncertainty, and let the drafter stay in control.