I believe vLLM has been the gold standard for inferencing but I'd like to learn how vLLM, langgraph, langfuse, other observability and business logic is meld into a single setup.