Worked on the kaito-project/kaito repository to enhance model deployment performance by implementing NVMe Local Caching for model files. This involved designing and integrating a caching layer that stores model files on local NVMe storage, which reduced model load times and inference startup latency. The approach included architectural changes, cache management, and prefetching strategies, all benchmarked to quantify performance improvements across deployment scenarios. Documentation was updated in Markdown to detail the new caching architecture and provide usage guidelines. The work focused on performance optimization and documentation, delivering measurable improvements to deployment speed and runtime responsiveness for model-serving workflows.