This paper presents a novel approach to legal judgment prediction by combining BERT embeddings with a Delaunay-based Graph Neural Network (GNN). Unlike inductive methods that classify legal documents independently, our transductive approach models the entire document set as a graph, capturing both contextual and relational information. This method significantly improves classification accuracy by enabling effective label propagation across connected documents. Evaluated on the Swiss-Judgment-Prediction (SJP) dataset, our model outperforms established baselines, including larger models with cross-lingual training and data augmentation techniques, while maintaining efficiency with minimal computational overhead.