Worked on extending model capabilities and enhancing transparency in machine learning infrastructure. In the volcengine/verl repository, enabled longer-context support for vLLM by allowing sequence lengths to exceed max_position_embeddings when RoPE scaling is configured, and supplemented this with user-facing documentation to guide configuration overrides. In kvcache-ai/sglang, implemented support for Mixture of Experts (MoE) routed experts in completions and chat endpoints, adding parameters and response schema changes to expose expert routing data for improved observability and debugging. Maintained backward compatibility and clear traceability throughout, leveraging Python, API development, and thorough documentation to deliver robust, maintainable backend features.