Worked on optimizing CUDA Docker builds and improving CI reliability across the jeejeelee/vllm and vllm-project/ci-infra repositories. Introduced a dedicated csrc-build stage in the Dockerfile to streamline CUDA kernel compilation, resulting in faster and more maintainable builds. Addressed Docker build cache invalidation in CI by conditionally passing VLLM_MERGE_BASE_COMMIT only when precompiled binaries were used, reducing unnecessary rebuilds and improving developer productivity. Leveraged Docker, Python, and YAML to enhance build automation and CI/CD workflows. The work demonstrated a practical approach to cross-repository collaboration, focusing on maintainability, build speed, and efficient use of caching strategies in complex environments.