The goal was to move from theory to a working classifier. I started with labeled handwriting data, normalized image sizes, and built a CNN in TensorFlow with convolution, pooling, and dense layers. Key steps: grayscale conversion, train/validation split, tracking accuracy and loss per epoch, and saving the model for inference on new images. The hardest part was balancing model depth vs. overfitting on a limited dataset. Data augmentation and early stopping helped stabilize results. This project maps directly to document-understanding workflows: preprocessing messy inputs, structuring features, and measuring accuracy on real samples.