Case study // Atlas Underwriting
Real-time credit inference on a 40ms risk budget
Moved credit decisioning from overnight scoring to inline inference at payment time, inside a hard 40ms latency envelope.
Approval rate
+340%
Inference p99
31ms
Stack
Triton · ONNX · Feature store · Go · FedNow
01 — Challenge
Atlas declined thin-file applicants because its gradient-boosted model ran in a nightly job against stale bureau data. Real-time scoring had been attempted twice and failed on tail latency.
02 — Architecture
Hot features are precomputed into a co-located store with 2ms reads; cold features fall back to a lighter model.
A latency-budget router degrades across three model tiers instead of timing out.
Every decision writes a signed explanation vector for adverse-action compliance.
03 — Outcome
Thin-file approvals rose 3.4× while charge-offs fell. The model now scores 9,200 decisions per second at peak with no batch jobs left in the path.