Financial fraud is constantly evolving as attackers adapt their strategies to evade existing detection systems. This creates a critical need for systems that can both adapt to new fraud patterns and respond quickly enough to prevent financial losses. To address this challenge, we are developing a system for Synchrony Financial that detects fraudulent consumer banking transactions using real-time streaming and advanced AI methods. Because low-latency, high-throughput performance is critical, the system follows a three-tier escalation architecture for the flagging of possible fraudulent transactions. The first tier is a rules engine that performs fast deterministic checks to flag transactions that violate predefined conditions, providing a computationally inexpensive first line of defense. Transactions that pass this stage are evaluated by the second tier, an unsupervised classical machine learning (ML) model, specifically an Isolation Forest, that performs robust anomaly detection. In high-dimensional data, able to inherently model complex feature interactions. The final tier consists of a deep learning (DL) model, a heterogeneous temporal graph neural network (HTGNN), which analyzes relationships and temporal patterns among entities such as transactions, cardholders, and accounts to identify more sophisticated and subtle fraud patterns that earlier tiers may miss. To support low-latency, high-throughput data ingestion from Synchrony’s systems, our platform leverages Kafka for distributed event streaming and Flink for real-time stream processing. The system also uses some cloud services from Amazon Web Services (AWS), providing scalable cloud infrastructure that seamlessly integrates with Synchrony’s enterprise ecosystem.