Very true, and it can be quite challenging to do it. E.g. if the ML model provides a score which is used to rank results, even estimating accuracy in production may not be straight forward. Looking forward to reading the article!