This study offers the best way to categorize legal documents in the Indian judiciary using DeBERTa, a transformer-based model known for its disentangled attention process. To tackle domain-specific problems like as data imbalance and overlapping terminology, we employ hyperparameter tweaking, class rebalancing using SMOTE, and advanced preprocessing. We accomplish this by using a large corpus of over 45,000 legal papers that are divided into five categories: family law, corporate law, criminal law, civil law, and constitutional law. Our trials beat baseline transformer models like BERT and RoBERTa, achieving 84% test accuracy and 84% macro F1-score. The suggested model shows great promise for automating the classification of legal documents, enhancing judicial efficiency, and establishing the foundation for more open and understandable artificial intelligence systems in the legal field.