• Built a PyTorch-like deep learning framework from scratch, implementing dynamic computation graphs and reverse-mode automatic differentiation.
• Supported core neural architectures (CNNs, RNNs, Transformers), optimizers, and GPU acceleration for end-to-end training.