Sentinel Pay: A Hybrid AI-Driven System for Real-Time Fraud Detection in Digital Payment Systems

Abstract

Financial system fraud goes undiscovered, even when it would be most beneficial for them to do so. UPI transactions typically complete in less than two seconds, making it nearly impossible to
detect any mistakes after the fact. SentinelPay circumvents this issue by classifying all transactions prior to the payment gateway approval at the authorization step. The architecture was developed using an event-driven microservices approach, and each part was designed to maintain the latency below acceptable boundaries. A few design choices shape the system. Apache Kafka takes care of transaction events to keep submission and scoring from being mixed up. There are various levels in the detection pipeline. After discovering statistical outliers and starting with the basic threshold criterion, Spring AI was used to make decisions based on the situation. There was no layer that was missing because they all depend on each other. This makes sure that the detection process works well and finds any mistakes. Redis may maintain user activity profiles, so the inference engine can acquire data in milliseconds. After looking at 50,000 entries in total, SentinelPay found real UPI fraud trends. We looked at the following numbers: recall (94.6%), detection accuracy (95.2%), F1-score (94.2%), and precision (93.8%). The throughput stays the same at 1,200 TPS, even when there is a lot of traffic. The average end-to-end latency is about 50 ms. This method gives results that are 13.2 points more accurate if you start with a rule. There are two other benefits: the frequency of false positives was down by 65%, and the time it took to get findings went down by 75%. You do not have to change your UPI settings to use SentinelPay.