A vision server that lets an AI read construction linework straight off PDF drawing sheets
Lets an AI actually see construction drawings, not just read their text.
The visual half of an automated millwork takeoff: it renders any drawing sheet or sub-region to a crisp image and returns it as model-readable content, so a vision model can count doors, drawers, shelves, and finished ends that are drawn, not labeled.
A look at the interface.
Representative UI — abstracted, never a client's real data.- Python
- MCP
- PyMuPDF
- Pillow
The industry's measurement tool returns text only, dimensions and callouts, and can't see drawn detail like a door-vs-drawer split or a finished end. Those distinctions are drawn, not written, so a separate server was needed to expose the visual half of a takeoff.
How it
works
The industry's measurement tool returns text only and can't see drawn detail like a door-vs-drawer split. This server renders any sheet or normalized sub-region to a crisp image and returns it as model-readable content. Find a sheet by full-text search, render it with a coordinate grid, then zoom a region at higher resolution to count detail. Every render is capped to a configurable pixel budget.
What I built
(03) Inside the buildNavigate-then-zoom workflow
Find a sheet by full-text search, render it with a coordinate grid for an overview, then zoom a normalized sub-region at higher resolution to count detail.
Exhaustive tile sweeps
Slice a sheet into high-resolution tiles for complete coverage when a single render would miss fine linework.
Megapixel budgeting
Every render is capped to a configurable pixel budget, auto-reducing resolution and reporting it instead of emitting a giant payload.
The
receipts
- Returns native image content so a vision model reads drawings directly
- Sheet addressing by page, sheet code, or text search
- Works on any PDF drawing set, with no proprietary software running
- Automatic resolution governance keeps payloads sane