Patent Figure Labeler

What it did

You uploaded a patent PDF of up to 50 MB, with no account. The browser turned each page into an image, and the server asked Google’s Gemini 2.5 Flash model whether the page was a drawing sheet and, if it was, where each reference numeral was on it. Cover pages and text pages were left unchanged.

For the element names, the server took the text layer of the PDF and asked the same model what each numeral found on the drawings referred to. It then drew a box reading “number: name” on the drawing sheet, in an empty area near the numeral where it would not cover the drawing or another label, or in the margin when there was no room. The result was a new PDF to download.

Both steps depended on the model. A numeral it missed on a drawing got no label, and the names were only as good as the model’s reading of the description, so the labels needed checking against the patent. The project’s code and notes contain no measurement of its accuracy.