Skip to main content
WZ SolutionsWZ SolutionsSoftware Engineering
Financial Services · 2025

Automating document intake with AI-assisted extraction

A retrieval and extraction pipeline that reads application documents, structures the data and routes uncertain cases to a human reviewer.

Client: Financial services provider (anonymised)

Document intake time, from tens of minutes per file
Minutes
Extraction output, validated before persistence
Schema-valid
Retained for low-confidence cases by design
Human review
Evaluation suite gating every prompt and model change
Versioned
The challenge

Applications arrived as scans and photographs of wildly varying quality. Staff re-typed the same fields into two systems, and errors surfaced only at approval, sending the file back to the start. Throughput was capped by how fast people could read.

What we built

We built an extraction pipeline combining OCR with a large language model constrained to a strict output schema, validated against business rules before anything is written. Confidence thresholds route ambiguous documents to a review queue with the source region highlighted. An evaluation set, established before development, is re-run on every model or prompt change so accuracy regressions fail the pipeline instead of reaching production.

Technologies used

  • Next.js
  • TypeScript
  • Python
  • PostgreSQL
  • Claude
  • Redis

Services engaged

More work

Other case studies

Recognise this problem?

If any of the above sounds like your operation, a short discovery conversation will tell us both quickly whether there is a fit.

Talk to usRequest a Quote

Or email walter.mussagy@gmail.com — we reply within one business day.