MEHDIOMehdio — practical data and AI education by Mehdi Ouazza

Jev for analytics might be one of its biggest use cases. We all have that one text column. Support tickets, reviews, complaints, an address nobody parsed. There's gold in there, but parsing it at scale gets expensive fast. Before, you had two options: - an LLM over every row: it works, but it's slow and it burns your credit card - a trained classifier: cheap at scale, but you need labelled data and a model to maintain Jev makes this easy and cheap. On MotherDuck, it's just a SQL function: prompt_jev(). You give it the text, a question, and the allowed answers. You get back a typed label and how sure it is. The combo I like most: an LLM designs the taxonomy once on a small sample, then Jev applies it to every row. On 100,000 consumer complaints: - gpt-5-mini proposed 7 labels from 40 rows in 7.6 s - Jev classified all 100k rows in 82 s, 0 NULLs - the rest is a plain GROUP BY I did a deep dive with the full SQL, real data, and how to check the labels before you trust them 👇 https://motherduck.com/blog/jev-for-analytics/