When people hear "AI-powered takeoffs," the mental image is usually something like a camera pointed at a drawing, trying to recognize shapes. That's one approach. It's not ours, and there's a reason.
Here's what actually happens when you upload a PDF to Vector Takeoff — in plain terms, without the sales pitch.
What's Actually Inside a PDF
A PDF isn't a picture. When an architect exports floor plans from their CAD software, the resulting PDF contains structured geometric data: lines, arcs, polygons, and text, each at specific coordinates.
That data is exact. The line that represents the north wall of the living room is stored as a vector path from point A to point B, with precise coordinates. The room label "Living Room" is stored as text at a specific position on the page. The dimension string "14'-6"" is text positioned to annotate a specific wall.
This is fundamentally different from a scanned PDF, which is just an image — pixels representing what the drawing looks like, with no embedded geometric structure. Scanned drawings are harder to work with accurately.
How We Read the Geometry
Our pipeline starts by parsing the PDF's vector layer. We extract every polygon, line, and text element and build a spatial model of the drawing.
Room boundaries are typically polylines or collections of line segments that form closed shapes. We identify these closed regions, associate them with nearby room labels, and compute their area and perimeter from the coordinate data.
This is the key step that makes measurement reliable. We're not estimating dimensions by looking at the drawing — we're computing them from the same coordinates the architect used when they drew it.
Where AI Vision Comes In
Not everything in a floor plan is simple geometry. Door and window symbols, for example, are standardized but vary between drafting styles. A door might be represented as an arc and a line, or as a block symbol from a CAD library.
This is where computer vision plays a role. We use a trained model to identify common drawing symbols — door swings, window casements, fixed windows, sliding doors — and map their positions to the rooms they serve.
The AI doesn't need to understand what a door is in any deep sense. It needs to reliably identify the symbol and locate it within the drawing's coordinate system. That's a classification and localization problem, and it's one that modern vision models handle well when trained on enough examples.
Cross-Reference and Confidence
No single data source is perfect. Dimension annotations sometimes conflict with scaled geometry. Room labels occasionally land inside the wrong polygon. Door counts can disagree between the floor plan and the door schedule.
Our pipeline cross-references multiple data sources and flags rooms where the sources don't agree. When we're uncertain, we tell you — so you can spend your review time in the right places rather than spot-checking everything equally.
What Accuracy Looks Like in Practice
In testing against projects that had been previously measured by hand, our pipeline produces outputs within roughly 10–15% of human measurements on well-formatted vector PDFs. That's close enough to be useful for estimating in most cases.
There are conditions where accuracy drops: unusual drawing styles, non-standard symbol sets, very complex geometry, or drawings that mix vector and raster content. We're transparent about these limitations, and we continue to improve.
The Honest Summary
The reason our approach works better than image-based recognition is simple: we're reading the same data your architect put in, not trying to infer it from a picture of the result. When the source data is high quality, the output is reliable.
That's the foundation the rest of the product is built on.