Built an end-to-end LLM pipeline from scratch in PyTorch, implementing core GPT/Transformer components including tokenization, embeddings, LayerNorm, GELU feed-forward networks, causal multi-head self-attention, Transformer blocks, and autoregressive text generation. Developed and trained GPT-style models, implemented temperature sampling and top-k decoding, and explored instruction fine-tuning using GPT-2 style architectures. Added checkpoint saving and loading for locally trained models. Built a Streamlit-based local inference interface that automatically detects model configurations from `.pth` checkpoints and provides configurable generation controls such as temperature, top-k, and maximum new tokens. The project demonstrates hands-on understanding of Transformer architecture, attention mechanisms, LLM training, instruction tuning, model inference, and practical deployment of locally trained language models. Tech: Python · PyTorch · Transformers · GPT · LLMs · Deep Learning · Streamlit · Model Fine-Tuning