Worked on performance and reliability improvements in distributed machine learning systems, focusing on two open-source repositories. In microsoft/DeepSpeed, addressed a multi-node logging issue by ensuring correct propagation of launcher arguments when using the PDSH runner, which improved observability and reduced debugging time for distributed training. In DarkLight1337/vllm, implemented a threaded logits processing pipeline in Python, introducing runtime configurability via an environment variable to optimize CPU-bound inference throughput. Leveraged skills in backend development, concurrent programming, and shell scripting to deliver targeted solutions that enhanced both runtime efficiency and operational reliability in complex, multi-node and multi-threaded environments.