Where they started
Halvard processes around 60,000 insurance claims a week. Intake ran on a queue, three worker services and a retry table that nobody fully trusted. When a worker died mid-claim, an engineer would find the half-finished record the next morning.
What they built
Each claim became one workflow run: extract, validate, check eligibility, price and file. Every call to a payer system is a step with retries and an idempotency key, so a timeout never files a claim twice.
Claims that fail validation now pause for a reviewer instead of landing in a dead-letter queue. The reviewer fixes the field in the dashboard and the run resumes at the next step.
What changed
The retry table and the queue are gone. On-call went from a weekly stuck-job page to none since March, and the team ships intake changes on Tuesdays instead of waiting for quiet weeks.
