Online payment fraud is a significant challenge for financial institutions worldwide, often resulting in considerable monetary loss and customer dissatisfaction. Traditional fraud detection systems operate on batch-processed data, leading to latency that delays fraud mitigation actions. This paper presents a scalable, real-time online payment fraud detection pipeline that leverages Apache Spark Structured Streaming and ONNX Runtime. Our implementation extends the framework proposed by Vanini et al. (2023), integrating anomaly detection, economic optimization, and risk forecasting within a streaming environment. We detail the system architecture, feature engineering, model integration, and decision logic, and provide extensive experimental evaluations demonstrating competitive fraud detection performance with millisecond-level latency. Our findings validate the feasibility and benefits of deploying online payment fraud detection systems capable of real-time response in production environments.