More than 2 billion people lack access to clean water, a major global issue that poses major health risks. Traditional techniques for assessing water quality are reliable, but are frequently time-consuming and labor intensive, causing a delay between testing and practical results. This work tackles the problem of unbalanced datasets in water quality assessment, where most samples are considered safe, leading to skewed machine learning (ML) models that cannot detect unsafe drinking water. We examine how machine learning (ML) approaches can be used to address this issue, concentrating on three models: k-Nearest Neighbors, an ensemble approach, and an XGBoost-SVM hybrid. To address class imbalance, we used the Synthetic Minority Over-sampling Technique (SMOTE) in conjunction with Edited Nearest Neighbor (ENN) filtering, as well as Recursive Feature Elimination (RFE) for feature selection, on a comprehensive dataset of 10,000 water samples with 12 physicochemical and 4 microbiological parameters. Our results show that the XGBoost-SVM hybrid model performed exceptionally well, achieving recall rates of over 85% for non-drinkable water samples while maintaining high precision. This study establishes a strong framework for real-time water quality monitoring and public health safety systems, emphasizing the importance of appropriate preprocessing and model selection in overcoming the challenges of imbalanced data.