Worked on stabilizing the MegaMoE data path and key-value management within the yhyang201/sglang repository, focusing on improving data correctness and reliability for top-k processing. Addressed a dtype mismatch between topk_ids and topk_weights, ensuring consistent data types throughout the data processing pipeline. Enhanced the FakeKVManager by adding missing kv_args, which prevented potential runtime errors and improved overall functionality. Utilized Python for development, applying skills in data processing and machine learning to increase production readiness. The work enabled safer experimentation and contributed to the robustness of the system, supporting both development and deployment environments for machine learning workflows.