Worked on performance optimization for the jeejeelee/vllm repository, focusing on improving model execution for AITER fused experts. Addressed inefficiencies by eliminating redundant copies in output buffers, which reduced memory overhead and accelerated inference for mixture-of-experts (MoE) workloads. The solution was implemented in Python and leveraged deep learning and machine learning expertise, specifically targeting performance bottlenecks without introducing dependencies on AITER-level changes. This approach enabled safer integration and streamlined the path for future enhancements. The work demonstrated a strong understanding of memory bandwidth constraints and contributed to more efficient MoE inference, reflecting depth in performance optimization within machine learning systems.