This study custom builds a convolutional neural network (CNN) to classify potato leaves into diseases such as Early Blight, Late Blight, and Healthy leaves. The model uses the PlantVillage dataset and preprocesses the images of leaves by resizing, rescaling, and augmenting data to make the model robust against changes in lighting, angles, and leaf condition. The CNN architecture consists of five convolutional blocks with ReLU activation, two max-pooling layers, two fully connected layers, and one softmax classifier at the end for multi-class classification. The model is fully evaluated using five-fold cross-validation and achieves an average accuracy of 97.4%. Other metrics such as accuracy, loss, confusion matrices, and ROC curves were provided, and the model proved to be effective. The results show the ability of the CNN to generalize well across different folds of the dataset, indicating its reliability as an automated diagnostic tool, and emphasizing the need for efficient AI solutions to enhance sustainable farming. This model tackles the issue of manual inspection by providing farmers with timely, automated, and accurate disease management that increases agricultural productivity and reduces crop loss.