Financial services
Sharing fraud-detection benchmarks across teams without sharing real transactions
Ledger replaced a frozen, real-transaction benchmark with a differentially-private synthetic dataset internal teams can iterate on freely.
0
real PII exposed to new teams
+2.1 pts
F1 on the locked real benchmark
10×
more engineers able to iterate
The challenge
Ledger's fraud team had a high-quality real benchmark, but legal restrictions meant fewer than ten engineers could ever see it. New hires waited months for data access, and external research partners were excluded entirely.
How they used Epineone
- Generated a synthetic transaction stream calibrated to match the real benchmark's distributional and temporal properties.
- Applied differential privacy guarantees (ε = 1.2) at the record level, validated by an independent reviewer.
- Published the synthetic benchmark internally so any engineer could prototype models without legal approval.
- Kept the real benchmark as the locked, final evaluation gate for production deployment.
"The synthetic benchmark unlocked our roadmap. People who'd been blocked for months were shipping fraud-model experiments in a week."