Small-then-large LLM preprocessing layer — like litellm.completion() with a smart routing preprocessor.