Loading…
Case study · Financial Services
A rules-based fraud engine that hadn't moved in five years was missing modern attack patterns and drowning the analyst team in false positives. We replaced it with a streaming ML pipeline on AWS — and stood up the AI governance frame the bank's regulators were starting to ask about.
The challenge
The bank's fraud engine was a five-year-old rules-based system that had accreted hundreds of hand-tuned thresholds. New attack patterns — synthetic identities, ATO via SIM-swap, mule-network laundering — were missing the rules entirely. False positives had crept above 5%, exhausting the analyst team and pushing legitimate customers into call centers.
Three constraints made this a hard project: (1) the bank's regulators had begun asking about model governance and the bank had no model-card or evaluation framework; (2) the rules engine was load-bearing and couldn't be turned off during cutover; (3) the analyst team's existing workflow tools — case management, alert triage, customer communication — had to keep working through the migration.
Our approach
We took a strangler-fig approach. Instead of replacing the rules engine in one cutover, we ran the new ML pipeline in shadow mode for three months, then in active-decision mode for low-risk segments, then migrated the high-risk segments last with a documented rollback path.
The new pipeline ran on AWS Kinesis + Lambda + SageMaker, with Redis for low-latency feature lookup. We co-designed the feature store with the bank's data team so the engineered features could be reused across other ML use cases (credit risk, marketing personalization) on the same substrate.
Model cards, evaluation harnesses, and audit logging shipped alongside the first model — not after. Every prediction logs the model version, feature snapshot, decision threshold, and downstream action. Regulators got the audit pull they needed in a one-day evidence run, not a six-week scramble.
The solution
The deployed system uses a two-tier ensemble: a fast linear scorer for sub-50ms decisions on the hot path, and a deeper gradient-boosted model for asynchronous secondary review. Both feed an analyst case-management UI we co-built with the bank's fraud ops team — preserving their muscle memory while surfacing the new model's reasoning.
Results
Six months after full cutover, the new system catches 37% more fraud than the legacy rules engine while cutting false positives by 42%. The analyst team's queue is back under control, and the bank's regulators completed an unscheduled exam in two weeks instead of the eight they'd budgeted.
In their words
“Prosigns didn't just replace our fraud engine — they gave us a model governance frame our regulators are happy with and our data team can run.”
Stack
Talk to us
A senior engineer plus the relevant department lead joins the first call. No discovery gauntlet, no junior reps.