Worked on the deepseek-ai/DeepEP repository to develop an experimental nvDev branch focused on reducing latency for small token workloads using CUDA Compute Fabric Transport. The approach involved implementing GPU programming techniques and leveraging CUDA to optimize transport mechanisms, specifically targeting performance improvements in token-constrained inference scenarios. Collaborated across teams to co-author and document the onboarding process for the experimental branch, enabling rapid evaluation of latency reduction strategies. The work established a foundation for future enhancements in low-latency GPU transport and contributed to the repository’s documentation using Markdown, ensuring that technical context and experimental results were clearly communicated for ongoing development.