Focused on model routing, caching, and cost-latency-quality tradeoffs.