(OSS) >90% of LLM calls don't need an LLM and could be replaced by classic ML (github.com)