← Back to work
AI & data · WEB APP · 2025
ParseFlow
Extraction you can trust, because a human signs off.
An AI workspace that reads submission packets, pulls the fields that matter, and routes only the uncertain ones to a person.
Client
Commercial insurance underwriter
Year
2025
Duration
4 months
Our role
Product, ML integration, review UX
THE CHALLENGE
Underwriters retyped data from hundred-page PDF submissions by hand. It was slow, error-prone, and the bottleneck grew every time the business did.
OUR APPROACH
- 01Built an extraction pipeline that returns every field with a confidence score and its source location.
- 02Designed a review UI where a human confirms low-confidence fields side-by-side with the document.
- 03Closed the loop — corrections feed back as examples so accuracy climbs over time.
WHAT HAPPENED
Turnaround on a submission dropped from hours to minutes, high-confidence fields flow through untouched, and reviewers spend their time only where the model is unsure.
- Faster turnaround
- 9x
- Fields auto-accepted
- 88%
- Extraction accuracy
- 99.1%
BUILT WITH
Next.jsPythonLLM APIsPostgreSQLRedis
IMPACT REPORT
Human-in-the-loop, measured
Extraction quality and reviewer load across the first 40,000 documents.
9× faster
hrs → minTurnaround
88%
high-confidenceFields auto-accepted
99.1%
was 96.2%Extraction accuracy
-84%
Reviewer touches
Extraction accuracy as corrections feed back
What happens to each field
Minutes to process one submission
GALLERY
