Mid-size correspondent

The problem

The team ran an in-house pricing engine that had grown organically over seven years. Investor-specific code paths were everywhere; every new investor required a deploy. Compliance audits routinely took three days to produce evidence because the per-rule trace was generated as an after-the-fact "explain pass" that frequently disagreed with the actual pricing run.

The breaking point was the lock-day surprise rate. Eligibility was a single-stage check that ran late in the pricing flow; the team regularly committed to a price that came back repriced or denied at lock — typically a dozen times per month, each one a $400–$600 ops incident.

The evaluation

Three vendors made it to the final round; two were eliminated when their per-rule trace turned out to be a separately-generated explanation that didn't survive the "does the trace match the price?" test on a synthetic edge case. RateStack's trace is built as a side-effect of the pricing math — verifying that against synthetic test cases convinced the architecture review.

The migration

Eight weeks from kick-off to go-live, with a three-week parallel run. Investor onboarding moved from "deploy a parser" to "upload a ratesheet and confirm a learned mapping template," and the team shipped two new investors in week six without engineering involvement.

The MISMO 3.4 / ULAD import endpoints replaced six bespoke ETL pipelines. The webhook DLQ replay tooling absorbed the operational surface that the in-house team had built (and kept maintaining) for retry-logic edge cases.

The results

Lock-day surprises fell from 12/month to fewer than 1/month — a 92% reduction tied directly to two-stage eligibility, which surfaces ineligibility reasons inline before the full ladder runs. Audit-response time dropped from 3 days to under 4 hours, driven by the per-rule trace plus historical replay. The engineering team redeployed two FTEs from pricing-engine maintenance to higher-leverage work.

What surprised them

• The mapping-template engine learned faster than expected — by month two, new ratesheets were onboarding without operator review on first arrival for previously-seen formats.
• The hash-chained audit log made the annual SOC 2 evidence collection "mechanical" (their word) — a single SQL extract instead of a multi-week reconstruction.
• Sandbox tier turned out to be useful for engineering on-call after go-live; the team kept a sandbox running for novel scenarios that they didn't want to test against the production tenant.