Developed and integrated a new NVSHMEM-backed communication backend into the facebookresearch/param repository, enabling faster intra-node and inter-node communication for PyTorch workloads. Leveraged CUDA and C++ to implement all-to-all communication patterns with conditional backend selection and memory symmetry handling, optimizing bandwidth utilization in multi-GPU environments. Addressed reliability by fixing unit test stability issues, ensuring continuous integration remained robust after backend changes. In the pytorch/pytorch repository, optimized inter-node communication by tuning thread block configurations based on GPU count, further improving performance for distributed deep learning. Demonstrated expertise in GPU programming, distributed systems, and high-performance computing using Python and C++.