PaySentry

Overview

Every fraudulent transaction that slips through costs money and erodes trust. PaySentry catches suspicious payments in under 100ms using ML-powered risk scoring, which is fast enough to block fraud before it goes through without holding up real customers.

Transactions flow through a layered pipeline that mixes traditional rule-based checks with ML anomaly detection. I built it to plug into live payment gateways where speed really matters, and it handles compliance reporting automatically so teams don't have to chase regulations manually.

Key Features

• Real-time transaction scoring with sub-100ms response times
• XGBoost fraud detection model trained on 500K+ synthetic transactions
• Automated compliance reports that adapt to different regulatory requirements
• Interactive analytics dashboard for visualizing fraud patterns and trends
• Redis caching layer for high-throughput transaction processing
• Configurable risk thresholds with multi-level alert escalation

How It's Built

PaySentry is structured as a layered microservice pipeline:

• Ingestion: FastAPI endpoints receive transaction payloads, validate schemas, and push events into a Celery task queue for asynchronous processing
• Feature Engine: Pandas and NumPy compute real-time features like rolling averages, velocity checks, and geo-distance anomalies from raw transaction data and cached user profiles in Redis
• Scoring: An XGBoost model trained on 500K+ synthetic transactions evaluates feature vectors and returns fraud probability with confidence intervals
• Decision Engine: Rule-based post-processing applies regulatory thresholds, triggers alerts, and generates compliance reports in PostgreSQL

Interesting Challenges

• Class Imbalance: Fraud makes up less than 1% of transactions. SMOTE oversampling combined with cost-sensitive learning boosts minority-class recall without sacrificing precision
• Cold-Start Problem: New users have no transaction history to score against. A fallback rule engine using device fingerprinting and geo-IP heuristics covers the gap until enough behavioral data accumulates
• Regulatory Complexity: Financial regulations vary by jurisdiction. A pluggable rules framework lets compliance logic be defined as configuration rather than hardcoded business logic

Screenshots