Worked on stabilizing build performance and reliability for the google-research/kauldron repository, focusing on environments with and without GPU support. Addressed a specific issue where GPU-related build flags, such as cuda_compress, were inadvertently affecting CPU-only builds, leading to unnecessary configuration noise and potential performance degradation. Implemented a targeted fix in Python, optimizing the build system to ensure that GPU-specific flags are applied exclusively when GPU accelerators are present. This change improved build reproducibility and continuous integration reliability by isolating GPU dependencies. The work demonstrated a strong understanding of build systems and Python development, emphasizing precise configuration management and environment-specific optimization.