Credit card fraud is a serious issue of this decade with increasing volume in digital transactions, leading to financial losses and security risks. Tracking fraudulent transactions is tough due to the highly imbalanced nature of transactional data and speedy fraud tactics. Real-time prediction with machine learning approaches seems to be one of the suitable solutions to avoid fraud in real time. This research applies machine learning models Random Forest, Decision Tree, XGBoost, CatBoost, LightGBM, and Artificial Neural Networks to identify fraudulent transactions from a highly imbalanced dataset. To make efficient model sensitivity, which is heavily skewed towards legitimate transactions is maintained using Synthetic Minority Oversampling Technique (SMOTE) and under sampling to improve model sensitivity. Every model is trained and tested individually and evaluated using various metrics such as precision, recall, F1-score, and AUC-ROC. In comparison to all models, XGBoost has the best performance, achieving the highest AUC-ROC score of 98.5% with a balanced precision and recall. It also has the lowest number of false positives and false negatives. Architecture is designed for low latency by using efficient use of optimized models and data pipelines. Once a model is trained and validated, the selected model is serialized and deployed through an API-based microservice capable of handling real-time transaction data.