Sentinel — Fraud Detection Platform
Production-grade fraud operations platform with calibrated LightGBM scoring at 8.5ms, SHAP explainability on every prediction, and $1.23M in modeled net savings from cost-aware threshold tuning.
Screenshots
A tour through the analyst workflow, MLOps surface, and admin configuration. Every screenshot is from the running app — no mockups, no Figma exports.
Fraud operations command center
The analyst's home base. Animated KPI cards, a live posture banner that surfaces elevated risk windows, risk mix bar, geographic distribution mapped from synthetic KYC enrichment, and a replay engine that streams synthetic transactions through the model in real time so the dashboard breathes with live data.
Login and brand surface
Split-screen entry with a fraud command center preview. Demo credentials are visible on the login card so recruiters can sign in without setup friction.
Fraud queue
The analyst worklist. Risk and decision filters, paginated, with real-time polling so newly scored transactions surface without a refresh.
Transaction detail with SHAP explainability
Every prediction comes with reasoning. The SHAP waterfall shows the top contributing features, and a plain-English analyst summary translates the math into language a fraud analyst can defend in a meeting. Decision buttons (confirmed_fraud, false_positive, escalated) record analyst feedback for future model retraining.
Entity profile
Account-level investigation. Counterparty network, risk trend, full transaction history, watchlist controls. Account IDs across the entire app deep-link to this view.
Multi-criteria investigation
Full-history search with stats strip, quick-select presets, bulk action bar, CSV export, and URL-synced filters for deep-linking.
Case management with SLA tracking
Cases bundle transactions, entities, and notes under a single investigation. Created from the queue, entity profile, or the investigate bulk action.
MLOps: drift monitoring and threshold tuning
Per-feature PSI tracking with baseline-vs-recent score distribution comparison. A threshold tuner that visualizes precision, recall, and net business savings against a configurable cost model — not just abstract ROC curves.
Hardened batch upload with audit trail
CSV ingestion pipeline hardened against six attack classes. Schema validation, 5 MB cap, per-user rate limiting, formula injection neutralization, role-based access, and a full audit panel.
Model registry and admin settings
Versioned model registry with stage badges and live test metrics. Admin-gated threshold control on the settings page alongside profile, tenant, and alert rule configuration.
Responsive design
Every page adapts from desktop to phone with off-canvas drawer navigation. Tables become horizontal scroll regions instead of squishing the layout.
