Building something with LLMs, RAG or CPU-efficient inference? Tell me about it.