15 Jul 2025· 2 min readHow we cut LLM inference cost by 35% without touching qualityModel routing, caching, and context compression — in the order they paid off. No heroics, mostly plumbing.#cost#llmops