Worked on enhancing reliability and developer experience across the dayshah/ray and Furion-cn/sglang repositories, focusing on backend and distributed systems challenges. Introduced thread-safe network parameter handling for the LightGBM trainer by removing global session dependencies and implementing per-session isolation using Python and locking mechanisms. Improved Ray DAG’s API robustness by refining argument validation and error messaging, and reduced log noise by suppressing unnecessary numpy warnings during tensor transfers. Authored documentation to support adoption of Ray Compiled Graph, highlighting performance and GPU communication benefits. Optimized scheduler efficiency in Furion-cn/sglang by eliminating busy waiting and improving process observability, leveraging system programming skills.