To address the limitations of latency and stale predictions in streaming financial surveillance and algorithmic trading systems, this paper introduces a low-latency, real-time feature and graph intelligence platform built using Apache Kafka, Bytewax, Feast, and Neo4j. The architecture integrates Apache Kafka for high-throughput ingestion of multi-source order, quote, and news streams, Bytewax for stateful stream processing and sliding-window aggregations, and Feast as a unified feature store with a Parquet-based offline data lake and a lightweight SQLite online store ensuring sub-millisecond feature retrieval. Furthermore, Neo4j is embedded to model multi-dimensional trade topologies (Account-Stock-Exchange relationships) for dynamic structural risk and PageRank analytics. We evaluate the platform using simulated high-frequency trading datasets, measuring real-time feature generation throughput, end-to-end pipeline latency, and anomaly detection performance under simulated market manipulation scenarios. A Spatio-Temporal Graph Neural Network (ST-GNN) anomaly detector trained on stream-computed features and graph embeddings achieved an F1-score of 0.962, significantly outperforming spatial-only (F1: 0.814) and temporal-only (F1: 0.785) baselines. Ablation studies isolating spatial graph connectivity and temporal sequence history confirmed that removing structural graph edges or temporal memory leads to sharp declines in predictive accuracy. Additionally, robustness testing under injected multi-account spoofing attacks demonstrated that real-time feature updates and graph correlation metrics effectively mitigate model performance decay. The proposed Kafka–Bytewax–Feast–Neo4j pipeline offers a scalable, production-ready framework for low-latency feature engineering and fraud-resistant market surveillance.