How would you architect a geo-aware system that queries multiple LLM APIs and normalizes their output?