Worked on enhancing Llama-based workflows in the pytorch/executorch and pytorch/ao repositories, focusing on input handling, evaluation, and model generation. Simplified prompt processing for Llama2, improved few-shot MMLU evaluation, and refactored token generation logic for better performance. Strengthened CI pipelines and testing frameworks, adding support for wiki and MMLU tasks and eager execution modes. Addressed reliability by fixing quantization and tensor initialization issues, ensuring stable model deployment. Leveraged Python, C++, and PyTorch to implement these features, emphasizing robust data evaluation, model export, and prompt engineering. The work demonstrated depth in both feature development and infrastructure reliability improvements.