Worked on the jeejeelee/vllm repository to enhance the reliability and maintainability of the FlashInfer CUTLASS Mixture-of-Experts (MoE) kernel. Addressed a critical bug by correcting the token bound parameter, which previously risked kernel misconfiguration during inference. The approach involved cleaning up unused configuration logic, streamlining the handling of kernel tuning parameters, and simplifying expert layer initialization and execution flow. These changes reduced the risk of misconfiguration and shortened debugging cycles, resulting in more robust MoE inference. Demonstrated strong backend development skills with a focus on CUDA, PyTorch, and Python, emphasizing kernel correctness and maintainable code structure.