Explainable AI In Court Case Analysis

The rapid growth of digital court records places a higher demand for automated tools that can analyze legal documents with consistency, accuracy, and transparency. This work presents a LegalBERT-based judgment prediction system that classifies court case documents into three verdict categories such as Positive, Neutral, and Negative, using real-world case texts collected from digitized PDF judgments and authorized legal repositories. A thorough preprocessing pipeline was followed to extract, clean, and standardize case facts, evidence summaries, witness statements, and argument sections to provide quality input for model training. The proposed system is fine-tuned on this curated dataset in a supervised learning manner that provides an academically reliable accuracy of approximately 95% on the held-out test set. It is integrated with an explainability layer, using SHAP and Integrated Gradients for token-level visual justifications of predictions. These explanations show the main legal factors, arguments, and evidence that influenced the decision of the model, thus providing transparency for legal and academic adoption. The system is deployed through a FastAPI backend with a user-friendly web interface that allows the classification of documents in real time, visualizing influential text segments. The results clearly show the capability of the proposed system to support legal research and decision assistance tasks by making fast, consistent, and interpretable verdict predictions. This study presents the first step toward robust AI-assisted judicial analytics and forms a basis for further work on integrating with more sophisticated legal reasoning approaches.