sentinel

Sentinel is an open-source, defense-only fraud detection system for cross-merchant identity reuse. It targets fraud rings that recycle the same UPI ID, phone number, or device fingerprint across multiple merchants, where each merchant individually sees a clean first-time customer. The project builds an identity link graph across customers, devices, VPAs, phones, and merchants, then scores each payment event with a deterministic weighted ensemble of seven published features. Verdicts are recommendations only: ALLOW, REVIEW, or BLOCK_REC, with reason codes and an evidence bundle. A separate LLM layer, using AWS Bedrock and gpt-oss-120b, only turns evidence into an analyst-facing narrative and never scores; it can be skipped if unavailable. The repository includes a FastAPI service, append-only SQLite audit store, and a Next.js analyst console called the watchroom. The console provides a ranked queue, cluster graph, taint path, signal decomposition, LLM narrative, evaluation dossier, and live ring replay. The README reports held-out test metrics such as precision 0.833, recall 0.882, F1 0.857, two of two rings caught, zero fraud silently allowed, and a net saving after false positives and review cost of +₹38,665 per 1,000 events. It also discloses that a GBDT baseline edged the rule ensemble on F1 and that slow-rate rings can evade current weights. Operational features include idempotent event ingestion, batch scoring, federated entity risk endpoints, graph cluster views with masking and admin-only unmasked access, append-only analyst feedback, Prometheus metrics, per-merchant JWT auth, champion/challenger shadow scoring, Docker Compose, and a zero-backend static demo snapshot. Quick-start commands cover setup, checks, evaluation, backfill, serving the API, running the console, container deployment, challenger training, load testing, and snapshot generation. All data is synthetic; no real PII is used, AWS keys are kept out of code and env files, gitleaks scans history, and verdicts never autonomously block or move money. Docs include a design suite, threat model, runbook, ADRs, and a what-broke log. Licensed MIT.