Document transcription
Scanned pages to formatted Word, with stamps and signatures kept in place.
- Client
- Client project
- Year
- 2026
- Status
- Live
- My role
- Sole engineer
Problem
Ordinary OCR reads the text on a scanned document and throws away everything else. For certified and legal work the stamp and the signature are the point, and losing them makes the output useless.
Architecture
- 01
Read each page
One Claude Vision call per page, run concurrently under a configurable ceiling, so a hundred-page file does not crawl through serially.
- 02
Find the non-text
OpenCV and scikit-image locate stamps, seals and signatures and record where on the page they sat.
- 03
Two output modes
Placeholder replaces each element with a text marker. Preserve extracts the image and re-embeds it at the detected position.
- 04
Async by default
Submit a job, poll its status, download through a presigned link. Storage is S3 or MinIO, with Alembic managing the schema.
Impact
Stamps and signatures survive the conversion instead of vanishing
Long documents process page-parallel rather than one at a time
Output is a formatted DOCX, ready to edit