Sah Shah worked on the pytorch/torchrec repository, focusing on enhancing training and evaluation pipelines over a two-month period. Using Python and leveraging skills in data processing and machine learning, Sah refactored the training pipeline to introduce reusable batch-processing methods, which improved maintainability and enabled seamless integration of latency tracking. In a subsequent project, Sah developed a hybrid training-evaluation workflow that introduced model state-based evaluation detection and robust data-draining logic, preventing evaluation batch loss and supporting reliable mode transitions. The work demonstrated thoughtful architectural improvements, addressing pipeline scalability and reliability while enabling better debugging and traceability within complex machine learning workflows.