This paper presents a unified, real-time framework for credit card fraud detection that integrates streaming machine learning inference, high-throughput data pipelines, and graph-based contextual analytics to uncover complex and coordinated fraud patterns that are often missed by traditional systems. Conventional fraud detection methods—typically reliant on static rules, threshold-based alerts, or offline batch machine learning models—struggle to capture the rapidly evolving behaviours characteristic of modern fraud schemes, which increasingly involve synthetic identities, automated attack patterns, shared digital infrastructure, and organized cybercrime networks. To address these challenges, the proposed architecture combines Apache Kafka for scalable and low-latency ingestion, Spark Structured Streaming for real-time feature engineering and ONNX Runtime–based machine learning scoring, and Neo4j for continuous graph modeling and anomaly detection across interconnected entities such as customers, devices, merchants, IP addresses, and financial accounts. By coupling model-generated transaction anomaly scores with graphderived relational insights—including shared device linkages, money mule substructures, coordinated merchant abuse, BIN-attack topologies, velocity bursts, geo-temporal inconsistencies, and other multi-hop behavioural signals—the system captures both individual transaction anomalies and higher-order fraud patterns that emerge only when entities are analyzed in relation to one another. This hybrid approach supports adaptive, event-driven risk scoring in which each new transaction can update feature states, graph embeddings, and network-level risk indicators in real time, enabling faster detection and response to emerging threats. The streaming ML layer ensures continuous, low-latency scoring that adapts to changing transaction characteristics, while the graph analytics layer provides structural context, community-based anomaly detection, and interpretable explanations derived from graph algorithms such as centrality, community detection, link prediction, and subgraph anomaly scoring. Together, these components form a cohesive fraud detection pipeline capable of high scalability, real-time reasoning, and continuous learning. Experiments conducted with synthetic transaction streams demonstrate that the integrated approach significantly improves early detection of coordinated fraud networks, reduces false negatives associated with relational fraud patterns, lowers detection latency, and enhances accuracy compared to traditional rule-based systems or standalone ML models. Moreover, the combined use of streaming ML and graph analytics produces richer explanations for flagged activity, improving transparency and aiding fraud analysts in rapid decision-making. Overall, the findings illustrate that merging high-throughput streaming technologies with graph-native fraud reasoning provides a powerful and adaptive solution for combating sophisticated financial fraud in modern, high-volume payment ecosystem.