Modern software systems generate massive volumes of log data and runtime metrics that contain early signals of impending failures. Reactive monitoring approaches often detect failures only after service disruption has occurred, leading to costly downtime. This paper presents an intelligent, end‑to‑end machine learning framework for proactive software failure prediction. The framework ingests heterogeneous log streams, extracts a rich feature set combining log event patterns, TF‑IDF vectors, and resource utilization metrics, and trains multiple classifiers Random Forest, XGBoost, and LSTM to forecast failures before they manifest. The system is evaluated on the public HDFS log dataset containing 11 million log entries, with failure instances labeled using system‑reported fatal events. Experimental results show that the Random Forest model achieves the best predictive performance with an accuracy of 96.2%, precision of 95.1%, recall of 91.3%, and an AUC‑ROC of 0.983. The LSTM model demonstrates the earliest lead time, detecting failure signatures up to 8 minutes ahead, albeit with slightly lower precision. A comprehensive feature importance analysis reveals that resource‑related metrics and specific log event counts are the strongest predictors. A discussion of false positive rates and deployment feasibility in real‑world continuous integration/continuous deployment (CI/CD) pipelines is provided. The findings confirm that a hybrid approach leveraging both structured metrics and unstructured log data significantly outperforms log‑only or metric‑only baselines, paving the way for self‑healing cloud applications.