Gas leaks pose significant safety and economic risks, and therefore, early detection is an imperative need in industrial as well as residential settings. Conventional gas sensors give reactive signals only after a significant leak occurs, which restricts them from preventing accidents. This work presents a machine learning predictive gas leak detection system deployed on an ESP32 microcontroller based on pressure pattern recognition. Pressure sensor readings are continuously being monitored and analyzed by a trained ML model that detects minor fluctuations and anomalies that signal the development of a leak. A hybrid of LSTM and XGBoost models is employed to detect both gradual and sudden leaks with improved accuracy. In contrast to threshold-based measures, the proposed system utilizes time-series pressure trends to predict possible severe declines before they reach critical points, thus offering an early warning system. Experimental results show that the system provides stable detection accuracy under a resource limited IoT platform. The proposed solution is scalable, low-cost, and holds great promise in improving safety in gas distribution networks, storage facilities, and domestic applications.