If we are speculating here, why not just go straight to an LLM serving all requests directly? No code needed.