This is predicated on you having training data already. I approach Jev more like Langchain -- you can prototype something new extremely fast and cheap, and if the use case works well enough, rip it out and build something bespoke. If it doesn't, you didn't spend a bunch of time curating a training dataset anyway.