Rate limiter and response cache

Sliding windows and LRU caches behind every AI API.