Developed and delivered the Distributed Semi-Sync Training Optimizer for the pytorch/torchrec repository, enabling semi-synchronous distributed training for large-scale machine learning models. This feature introduced a structured approach to local and global optimization steps, improving both scalability and training efficiency in distributed systems. The implementation leveraged Python and PyTorch, focusing on distributed systems concepts and advanced optimizer design. By allowing local and global updates to be coordinated in a semi-synchronous manner, the work provided a foundation for more predictable convergence in distributed training environments. The contribution addressed the need for efficient, scalable optimization strategies in modern machine learning workflows.