What it is
Training-free fraud triage with typed answers and an auditable policy.
What it does: Enriches each order with 38 signals, masks personal data, asks Jev four typed questions in one request and maps them to an action with reasons; rules keep checkout running if Jev is down.
Why it belongs in the showcase: It benchmarks Jev honestly against rules and a trained model, including the cases where Jev loses.
Caveats: Synthetic data from one author; no license file at intake.
Evidence
Public repo with a TypeScript scoring API, a Python simulator and a benchmark. On 44,097 held-out simulated transactions, Jev without training data caught 75.3% / 81.3% / 89.9% of fraud when flagging the top 2% / 3% / 5% of orders, versus 92.6% / 93.7% / 95.0% for XGBoost trained on 89k labelled orders from the same generator; Jev plus rules reached 90.4% at 5%. Median latency 127 ms, p99 248 ms, zero errors across 3,761 calls. Jev never sees names, emails, card data, IPs or addresses; a block needs both a high probability and a confident block recommendation.
Caveats
- All data is synthetic, and the simulator, features and rules share one author
- No license file in the repo at intake
- Jev thresholds were not calibrated, so it adds more 3-D Secure step-ups than XGBoost (2.3% vs 1.05% of legitimate orders)
- The Docker Compose stack has not been run end to end
Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.