Builds a high-performance storage engine for LLM inference and GPU training, accelerating massive KV-cache offloading and small-file reads and writes.