Worked on enhancing experiment tracking and reproducibility for large-scale deep learning projects in the Megatron-LM repositories, focusing on both swiss-ai/Megatron-LM and ROCm/Megatron-LM. Developed and integrated Weights & Biases (wandb) artifact tracking for model checkpoints using Python, introducing utilities and callbacks to automate artifact logging and loading. This approach established a robust ML Ops foundation, enabling seamless experiment comparison and improved auditability across distributed systems. By extending wandb_utils.py and implementing checkpoint callbacks, the work facilitated end-to-end visibility into training runs, supporting better collaboration and faster iteration for deep learning workflows without introducing new bugs during the development period.