Worked on enhancing distributed training workflows in the pytorch/torchrec repository by developing a GeneralizedCommsBandwidth class, enabling users to estimate bandwidth based on hardware and network configurations for improved resource planning. Focused on performance optimization and reliability, addressed two high-impact bugs: one corrected forward and backward communication estimations in sharding performance estimators, and another fixed import path resolution in the facebookresearch/param ET Replay tool. Leveraged Python, distributed systems expertise, and unit testing to ensure robust solutions. The work emphasized accurate performance estimation and tooling stability, supporting large-scale training and inference scenarios with improved planning and operational reliability.