We use it in production. Tested a bunch of alternatives (openai decision api, llm based approaches) and they are either too slow or not as good. But mostly too slow.