During September 2025, Sanjana focused on stabilizing and validating learning rate scheduling for large-scale training in the NVIDIA-NeMo/Megatron-Bridge repository. She addressed a critical bug by correcting the warmup calculation to use total decay iterations multiplied by the global batch size, ensuring learning rate schedules accurately reflect intended training dynamics. Using Python and leveraging her skills in configuration management, deep learning, and testing, she updated unit tests and configuration logic to align with the revised approach. This work improved training stability and convergence reliability, reducing the risk of mis-scheduled learning rates and supporting more consistent experimental outcomes across machine learning workflows.