Every drawing has a title block. Let AI read it.

What a title block holds

• Drawing number
• Revision
• Date
• Author / drafter
• Scale
• Client
• Project
• Address

Every supplier lays out those fields differently, with different labels and often in another language.

The approach

First find, then read

Have a model find the title block on the sheet and crop it before it reads the fields. On a full A0 sheet the small text in the title block gets lost otherwise.

Run it locally

Open vision models run on your own hardware, even on a server without internet access. Then no drawing leaves the organisation, which makes the GDPR question a lot simpler.

Scans, PDF and DWG

A vision model looks at the image, so a scan needs no text layer. PDF and DWG are converted to an image first. Faded scans remain the hardest.

Understand, not just read

Plain OCR turns the image into loose text and doesn't know what belongs together. A vision model understands what it sees: it links each label to the right value, knows that Tekening nr, Drawing Nr and DRG NO mean the same thing, and sees that a bare 'Nr' in a revision block is the revision, not the drawing number.

Open output

Deliver the fields as CSV or JSON. You can then load them into your document management or records system and link them to the right file, without replacing anything.

Measure, don't guess

Keep a set of checked examples and score every change per field. The same model doesn't always give the same answer on the same drawing, so one successful test says little.