Built a 2-layer feedforward classifier for MNIST (784→10→10) using only NumPy and Pandas, no ML frameworks. Implemented forward propagation, ReLU, softmax, cross-entropy loss, backpropagation, and full-batch gradient descent from first principles. Trained on 59,000 images with manual matrix operations, reaching ~80%+ training accuracy.

Built a customized Arch Linux desktop environment using Hyprland and Rofi, configured through modular dotfiles for reproducible system setup. Optimized workflow efficiency with a minimal, keyboard-driven interface and reduced UI overhead through lightweight compositing and window management configuration.