I reduced LLM inference GPU calls by 94% using semantic routing — Other | indexator.ai | Indexator.ai