Worked on distributed deep learning infrastructure, focusing on stability and reproducibility in large-scale training environments. In the volcengine/verl repository, implemented deterministic RANK ordering for distributed training checkpoint resumes by introducing a node IP-based sorting mechanism within RayWorkerGroup, ensuring consistent RANK assignment and reliable recovery of sharded model and optimizer states. Later, addressed robustness in the vllm-project/vllm repository by fixing expert_map handling in the FusedMoE layer, registering it as a named buffer to prevent misalignment during wake and sleep cycles. Leveraged Python, Ray, and PyTorch, demonstrating careful attention to distributed systems, checkpointing, and model optimization challenges.