Skip to content
Healthcare · North America

Automated claims intake and reclaimed 20 hours a week

Four coordinators spent their mornings retyping referral documents into the practice management system. We built a document pipeline that extracts, validates, and files them — with a review queue for anything it is not sure about.

Client
Multi-site healthcare network
Duration
7 weeks
Team
2 engineers, 1 architect
Delivered
2025

Measured impact

20 hrs
Reclaimed weekly
across four coordinators
99.2%
Field-level accuracy
measured against audit sample
18 min
Intake to system
down from 26 hours
94%
Straight-through rate
no human touch required
01 / Challenge

What was actually wrong

Referrals arrived as faxes, scanned PDFs, and email attachments in dozens of layouts. Coordinators keyed them in by hand, which was slow, error-prone, and — because entry lagged intake by a day or more — pushed scheduling and billing behind as well.

  • About 1,100 documents a month across 14 referring practices
  • 20 hours a week of manual data entry across four staff
  • 3.4% transcription error rate reaching the billing system
  • Average 26-hour lag between document arrival and system entry
02 / Engineering

What we built and why

The design principle was that the system should be trusted exactly as far as it deserves. Every extraction carries a confidence score, and anything uncertain goes to a human rather than into the database.

Stack
Next.jsPythonPostgreSQLn8nClaudeDockerSentry
01

Layout-aware extraction with strict schemas

Documents are parsed with layout awareness, then extracted against a strict schema. Invalid output triggers a repair pass; a second failure routes to review instead of guessing.

02

Validation against real-world rules

Insurance identifiers, dates of birth, and procedure codes are checked against reference data and internal records before anything is written, catching errors the model cannot see.

03

A review queue people actually like

Low-confidence fields surface side by side with the highlighted source region. Correcting one is two keystrokes, and corrections are logged as evaluation data.

04

Filed automatically, monitored continuously

Approved records write into the practice management system through its API with idempotency keys, and a dashboard shows throughput, confidence distribution, and failures in real time.

The review queue is what sold the team on it. Nothing goes in unless it is certain, and when it is not certain a person decides. That is the only version of this our compliance lead would have approved.
Director of Operations · Multi-site healthcare network

Let us look at your architecture.

Two ways to start, both of them short. Bring the problem, not a specification — the first useful thing we do is tell you what we would build and roughly what it costs.

Book an architecture call

Fifteen minutes, no deck. We map your problem to an approach and tell you what a realistic scope and budget look like.

  • A specific technical recommendation
  • A budget band you can plan against
  • An honest answer if we are the wrong fit

Send a written brief

Prefer to write it down? Email us the shape of the problem and we will reply with a first take, usually under 12 hours.

  • Goes straight to an engineer, not a sales inbox
  • We reply with an approach, not a brochure
  • Attach anything: repos, docs, screenshots