Reading archive documents automatically
Scanned documents are read, sorted by type and their key fields pulled out; an operator checks only the doubtful places.
- Client
- A contractor digitising a state archive
- Industry
- Public-sector contractors
- Went live
- July 2025
- Duration
- 6 months
- Team
- 6 people
What was there
The archive held documents from the 1950s on, in Azerbaijani written in both Cyrillic and Latin script. Operators typed every file by hand, and on that plan digitising would have taken 9 years.
What we built
- OCR trained for Azerbaijani in Cyrillic and Latin script
- Document type detection: applications, orders, certificates, acts and others
- Extraction of fields such as names, dates and numbers, with a confidence level
- A workstation where the operator checks only doubtful fields
- Everything on the client’s servers, with no internet access
Results
2.4m
pages read
2025
−71%
manual entry time per file
2025
96.4%
of fields read correctly after checks
test set of 12,000 pages
Operators no longer type, they check. We cut the plan from 9 years to 3.
Head of the digitisation project, a contractor digitising a state archive
Technologies
- Python
- PyTorch
- Tesseract OCR
- PostgreSQL
