Developed a concurrency-optimized prompt caching and sampling stack for the mudler/LocalAI repository, focusing on backend development and data structures using Python. The work introduced a thread-safe LRU prompt cache with per-request isolation, addressing race conditions in concurrent environments and improving throughput and reliability. Enhanced the MLX LLM pipeline by adding min_p and top_k sampling support, enabling more flexible sampling strategies. Implemented a trie-based prefix matching mechanism for cache reuse, with LRU eviction and per-entry key-value state management. Comprehensive unit tests were written to validate thread safety, cache behavior, and sampling logic, ensuring maintainability and robust production performance.