Tomato leaf diseases often show overlapping symptoms at early stages, making timely identification challenging for farmers. In this study, we make use of a DenseNet121 model using transfer learning for multi-class tomato leaf disease classification. The model was trained on the PlantVillage dataset using a Google Colab T4 GPU environment, and image augmentation was applied to improve robustness against lighting and background variations. The proposed system achieved a validation accuracy of 97.5% with balanced precision and recall across 13 disease categories, including visually similar ones such as Early Blight and Target Spot. To support expert validation, principal component analysis (PCA) was used to visualize the latent feature distribution and assess class separability. Since many real-world agricultural settings rely on smartphones or low-power devices, the model was further optimized for near real-time inference on edge hardware. This work demonstrates a useful study in precision agriculture, helping reduce crop losses through timely decision-making.