Fraud Detection for Online Banking using TigerGraph and Vertex AI

Challenge

The bank needed to analyze 2+ TB of application and transaction data to uncover hidden relationships among devices, addresses, IPs, merchants, and cardholders. However, traditional relational databases and commercial graph solutions couldn’t process this volume fast enough to meet sub-second detection requirements. As a result, the existing system lacked real-time fraud prediction and explainability, limiting the bank’s ability to identify and respond to emerging fraud patterns in time.

Project Scope

To combat the growing sophistication of fraud schemes, the bank required a solution capable of analyzing massive, interconnected data in real-time and providing explainable, actionable insights. Our team developed a graph-based fraud detection ecosystem that combined the power of TigerGraph, Vertex AI, and scalable cloud infrastructure.

The goal was to enable the bank to detect hidden fraud rings, assign risk scores within milliseconds, and empower compliance teams with transparent, data-driven decision tools.

Step 1 – Graph-Based Relationship Modeling with TigerGraph

We designed a graph schema linking users, cards, IPs, devices, addresses, and merchants, enabling multi-hop queries to uncover indirect links and detect fraud rings. TigerGraph delivered sub-second responses on large-scale data.

Step 2 – Real-Time Machine Learning with Vertex AI

We used TigerGraph features, combined with transactional data, to train fraud detection models on Vertex AI. These models identified anomalies and assigned risk scores in real-time. The models were continuously retrained on fresh data to stay ahead of fraud evolution.

Step 3 – Seamless Integration into Existing Infrastructure

Our solution was integrated with the bank’s existing Java-based services and Kafka-based transaction streams. As new transactions were ingested, relevant data was updated in TigerGraph, features were extracted, and the fraud score was fetched from Vertex AI.

We used a gradient boosting model (based on XGBoost) trained on historical fraud patterns, graph-based features (e.g. number of shared connections, abnormal link structures), and transactional behavior. This model was optimized for high precision to minimize false positives while ensuring timely fraud detection.

Detection results were then stored in PostgreSQL and used to trigger further actions like flagging transactions, alerting the fraud team, or initiating card holds.

Step 4 – Dashboards and Alerts for Compliance and Operations

• To support the bank’s fraud and compliance teams, we built interactive dashboards that displayed real-time alerts, fraud risk scores, and visualized connections between suspicious entities.
• A React frontend allowed analysts to explore fraud cases, view linked users or devices in a graph view, and manage alert statuses.
• Fraud detection results were made available through internal APIs, enabling seamless integration with internal tools and case management systems.
• This part of the solution ensured that both business and technical users could take timely action based on AI-driven insights and graph-powered visualizations.

Results and Achievements

The client achieved significant gains in detection performance, latency reduction, and fraud prevention. Highlights include:

• Detection accuracy improved by 38%, enabling earlier intervention
• Graph traversal reduced from minutes to milliseconds, thanks to TigerGraph
• 5x faster response time for fraud checks (<250ms total pipeline time)
• Near real-time blocking of suspicious cards across channels
• Enhanced compliance reporting with traceable graph paths for flagged cases