Sentinel AI — Fraud Detection Co-Pilot
The first AI Co-Pilot that stops fraud and explains why. Real-time fraud detection with explainable AI for financial institutions.
Overview
Production-grade AI fraud detection platform deployed for financial institutions, lending platforms, and fintech companies. Sentinel AI acts as an intelligent ride-along partner for fraud review teams, providing real-time fraud probability scores (0.0–1.0) and plain-English explanations powered by GenAI.
Built at Reallytics.ai for Tower Loan and other financial services clients.
Architecture
┌──────────────────────────────────────────────────────┐ │ Data Integration Layer │ │ OLL/TLOS Loan Application Systems │ │ - Secure data ingestion │ │ - Encrypted PII handling │ │ - ~100 raw data points per application │ └─────────────────────────┬────────────────────────────┘ │ ┌─────────────────────────▼────────────────────────────┐ │ Feature Engineering Pipeline │ │ - 650+ predictive features generated │ │ - Behavioral anomaly detection (deltaH timing) │ │ - Contact information pattern analysis │ │ - Profile stability scoring │ │ - Identity verification signals │ └─────────────────────────┬────────────────────────────┘ │ ┌───────────────┼───────────────┐ │ │ ┌─────────▼──────────┐ ┌────────────────▼───────────┐ │ XGBoost Detective │ │ Isolation Forest Watchdog │ │ (Supervised) │ │ (Unsupervised) │ │ - Known fraud │ │ - Novel fraud patterns │ │ patterns │ │ - Mathematical anomaly │ │ - 25% recall │ │ profiling │ └─────────┬──────────┘ └────────────────┬───────────┘ │ │ ┌─────────▼───────────────────────────────▼───────────┐ │ Ensemble Task Force (50/50) │ │ Combined score → 50% fraud detection on holdout │ └─────────────────────────┬───────────────────────────┘ │ ┌─────────────────────────▼───────────────────────────┐ │ Persona Classification (UMAP + HDBSCAN) │ │ - Digital Ghost (70% fraud concentration) │ │ - High-Friction Anomaly (abnormally slow process) │ │ - Safe Bet (100% legitimate, fast-track) │ └─────────────────────────┬───────────────────────────┘ │ ┌─────────────────────────▼───────────────────────────┐ │ GenAI Explanation Engine (Amazon Bedrock) │ │ - SHAP value interpretation │ │ - Plain-English PDF reports │ │ - Risk factors + mitigating factors │ └─────────────────────────────────────────────────────┘
Key Features
- Ensemble Detection: Task Force combining XGBoost (supervised, known patterns) + Isolation Forest (unsupervised, novel fraud) — 50% fraud detection on holdout
- 650+ Engineered Features: Behavioral anomalies, timing patterns, contact signals, identity verification, profile stability
- 3 Applicant Personas: Unsupervised UMAP + HDBSCAN reveals Digital Ghost (70% fraud), High-Friction Anomaly, and Safe Bet personas
- Explainable AI: GenAI-powered PDF reports via Amazon Bedrock translating SHAP values into plain English
- Real-Time Scoring: Headless API on AWS Lambda/SageMaker with API Gateway — scores applications at pre-funding stage
- Fraud Indicators: Detects early reversals, legal actions, repos, UCC failures, forgeries, first-payment defaults
- Continuous Learning: Automated retraining pipelines with data drift detection and A/B model deployment
Tech Stack
| Category | Technologies |
|---|---|
| ML Models | XGBoost, Isolation Forest, UMAP, HDBSCAN |
| Feature Engineering | Python, Pandas, NumPy, scikit-learn |
| Explainability | SHAP, Amazon Bedrock (GenAI reports) |
| Cloud | AWS Lambda, SageMaker, API Gateway, S3 |
| Data | PostgreSQL, encrypted PII handling |
| MLOps | CloudWatch, QuickSight, automated retraining |
| API | FastAPI, REST endpoints |
Results
| Metric | Value |
|---|---|
| Fraud detection rate (holdout) | 50% |
| Features engineered | 650+ |
| Applicant personas discovered | 3 |
| XGBoost recall | 25% |
| Ensemble improvement | 2x over single model |
| Scoring latency | < 500ms |
| False positive rate | < 15% |
Industries Served
- Financial Institutions
- Lending Platforms
- Credit Unions
- Online Lenders
- Fintech Companies
Source Code: The production source code for this project is maintained in a private repository due to proprietary and client confidentiality requirements. This repository documents the architecture, design decisions, and technical approach. For code-level discussions or collaboration inquiries, feel free to reach out.
Author
Rehan Malik — Senior AI/ML Engineer @ Reallytics.ai