Back to work

Client
US healthcare payer
Role
Owner of discovery, architecture, delivery, and documentation
Timeline
Discovery through production rollout
Team
Solo owner across the full lifecycle
Stack
  • AWS Bedrock
  • Bedrock Flows
  • LangChain
  • Python
  • Docker
  • AWS Lambda
  • S3
  • Secrets Manager

Automating Medicare claim payments at 95 percent accuracy

95%
UAT accuracy
20–30
batches per day

A US healthcare payer processed EDI 835 claim payments by hand. An AWS Bedrock pipeline with LangChain and Bedrock Flows now handles 20 to 30 batches a day, UAT validated at 95 percent accuracy across the enabled claim types, with a feedback loop that folds user corrections back in.

Problem

A US healthcare payer processed EDI 835 claim payments by hand. Claim payment reconciliation is regulated, error prone, and unglamorous work, and doing it manually put a hard ceiling on volume. The work also has no tolerance for quiet errors, so any automation had to be measurably accurate before anyone would switch to it.

Approach

I owned this one end to end: discovery, architecture, delivery, and documentation. The pipeline parses each EDI 835 file, extracts payment detail through AWS Bedrock using LangChain and Bedrock Flows, then validates the result before anything is posted. Whatever fails validation goes to an exception queue for a human rather than through on a guess.

Architecture

A cloud native Python pipeline on Docker and AWS Lambda, with S3 for staged files and Secrets Manager for credentials, wrapped around a Bedrock extraction flow with validation and an exception queue on the far side.

The hard part

Accuracy matters here more than model choice, so most of the effort went into validation and into proving the numbers in UAT rather than into prompt tuning. Production accuracy also drifts, which is why user corrections are captured as they happen and re-incorporated into the pipeline instead of being logged and forgotten. The correction loop is the part that keeps the ninety-five percent figure from being a launch-day number.

Outcome

UAT validated at ninety-five percent accuracy across the claim types the client enabled. The current rollout processes twenty to thirty batches per day, with corrections from real use feeding back in. Discovery, architecture, delivery, and the documentation behind it were all mine.

Next

Sole architect on a ten track lending and mortgage accelerator