How would you architect for both real-time chat and async batch inference on the same platform at OpenAI?
