Guardian ML — Real-Time Fraud Detection & Graph Analysis
The Problem
System Bottlenecks & Technical Friction
High-throughput payment processors face sophisticated fraud patterns that bypass traditional static rule-based filters. Rules-based detection produces high false positive rates (often >10%), frustrating legitimate users and driving checkout abandonment. Batch ML models retrained on weekly schedules lag emerging fraud rings by several days. Payment networks need sub-100ms real-time ML inference capable of graph entity resolution across cards, IPs, devices, and phone numbers without adding checkout friction.
Our Engineering Approach
Architecture Design & Implementation
Guardian ML is a dual-model fraud detection architecture combining gradient-boosted trees with real-time graph network analysis. Incoming transactions trigger 180+ low-latency feature calculations in Redis. An XGBoost model scores transaction anomalies in <30ms, while Neo4j executes sub-graph traversals to identify coordinated attack rings sharing device or identity fingerprints. Apache Flink streams transaction events, executes feature extraction, and returns decision payloads with explanation codes in 79ms P95 latency.
Technical Architecture
System breakdown & stack.
XGBoost gradient-boosted tree ensemble trained on 24M labeled transactions with SMOTE class balancing
Neo4j graph database modeling entity relationships across 8.7M cards, devices, IPs, emails, phone numbers
Apache Flink stream processing: Kafka ingestion → feature extraction → model inference → Neo4j graph query
Redis feature store with <5ms P99 read latency for 180+ precomputed transaction features
Cassandra time-series storage for 90-day rolling transaction history and behavioral analytics
FastAPI decision API with circuit breakers, fallback to rule-based logic on model service failure
Prometheus + Grafana observability: P50/P95/P99 latency, fraud recall, false positive rate, throughput
React ops dashboard with real-time fraud ring visualization, transaction drill-down, manual review queue
Engagement Timeline
Engineering delivery schedule.
Historical data extraction, EDA, feature engineering research, SMOTE balancing pipeline
XGBoost hyperparameter tuning (Optuna), cross-validation, threshold calibration for precision/recall
Neo4j graph schema design, entity resolution logic, Cypher query optimization, fraud ring detection algorithms
Apache Flink stream processing pipeline, Kafka integration, Redis feature store build
FastAPI decision API, circuit breakers, fallback logic, Prometheus instrumentation, React ops dashboard
Shadow mode parallel-run validation, latency load testing, production cutover
More Case Studies
Related AI projects
Want results like these for your business?
Book a 30-minute scoping call with a senior engineer. We'll scope your AI project, define the architecture, and give you a fixed-price proposal within 5 business days.