Developed a custom 211,406 parameter autoregressive language model from scratch using PyTorch, implementing core components like multi-head self-attention, positional embeddings, and layer normalization. Built a character-level tokenizer, custom training loop, and data pipeline using NumPy and PyTorch with zero external NLP libraries.