Batch claims processing is one of those systems that works until it does not — until a member is waiting, an operator is refreshing a screen, and everyone realizes that “overnight” is a euphemism for too late.
The goal was not novelty. It was reality in real time — without taking down the plane while we rebuilt the engine.
The constraint everyone felt
Mission-critical healthcare operations do not get maintenance windows with applause. Zero downtime was not a slide-deck ambition. It was the price of admission. Every migration step needed a rollback story that did not depend on heroics at 2 a.m.
The approach
We designed Kafka-based event pipelines with the unglamorous bones that keep systems standing: idempotent consumers, schema governance, dead-letter handling, and observability hooks that operators could actually trust.
Migration was incremental — strangler patterns, dual-write phases, automated rollback paths — so we could prove correctness before we cut over. The nervous system had to keep signaling while we rewired it.
What it unlocked
Latency dropped from hours to seconds. That alone changed how teams thought about operations — decisions closer to the moment reality changed.
But the compounding return came later. The same event backbone that modernized claims became the trusted feed for AI-powered claims assistance. We did not build the backbone for AI. AI could not have been built without it.
That is platform work at its best: invisible when it works, indispensable when the next workload arrives impatient and loud.