Loan applications read themselves
Document intake dropped from twenty minutes to under two.
- manual data entry
- -89% manual data entry
- field extraction accuracy
- 97% field extraction accuracy
The problem
Every application arrived as photographed documents. Officers retyped the same fields into three screens, and typos surfaced weeks later at signing.
What we did
- 01
Built an extraction pipeline over a language model, scoped to the six document types that make up most of the volume.
- 02
Kept a human in the loop: every extracted field shows its source crop and can be corrected in one click.
- 03
Measured accuracy against a labelled set before letting the pipeline touch live applications.
Where it landed
Officers review instead of retype. Errors are caught at intake, and the evaluation set keeps the model honest after each change.
Built with
- Next.js
- Python
- PostgreSQL
- Claude API
Tell us what is not working.
A short description of the problem is enough to start. We reply within one working day, and the first conversation is free.