With the rapid expansion of digital payment systems in India, the detection of fraud has emerged as one of the key requirements to ensure secure financial transactions. In this context, we propose a hybrid fraud detection architecture that utilizes domain-driven, rule-based heuristics in conjunction with machine learning classifiers to strike an appropriate balance between interpretability and detection performance. It is initiated with a transparent risk-scoring layer that classifies transactions into the Legitimate, Suspicious, and Fraudulent categories based on empirically defined thresholds on transaction amount, change in device or IP, temporal behavior, and merchant risk. Those rulegenerated signals are used afterward in a supervised learning pipeline that comprises Random Forest, SVM, and a hybridkernel SVM combining linear, polynomial, and RBF kernels. To handle the severe class imbalance inherent in financial fraud datasets, we have applied SMOTE-based oversampling during model training. The system is trained and tested on 1,000 realworld, labeled Indian digital transactions. The experimental results obtained indicate that the standard SVM has performed best, with a balanced accuracy of 84.1%, demonstrating a strong capability for minority class detection. It points out the highest overall accuracy to be due to the Random Forest model at 94.5%. The statistical validation performed with McNemar’s test identifies a statistically significant performance difference between standard SVM and the proposed hybrid-kernel SVM. It is expected that the proposed hybrid architecture provides an effective tradeoff between explainability, adaptability, and accuracy, hence finding wide practical applicability in the realworld banking and financial domain, where transparency and reliability are crucial features of any model or technique employed.