A different AI toolbox, for people interested in results, not hot stuff.
My pet theory is that the popular libraries and frameworks we have today (LangChain, LlamaIndex) are first generation products, where just getting the damn thing to work is less important than developer experience, and it’s not yet obvious what the code patterns will be. These are the products built by, and used by, the super early adopters, and leave a lot to be desired.
This is not to denigrate either them or the effort and skill people put into them! But I’ve got a “there must be a better way” feeling about the whole thing. Contrast the myriad web frameworks with wild ideas before some sort of best practices coalesced around Rails, Django, Laravel etc (or the equiv in JS land 10 years later).
This thinking is heavily influenced by Marvin ( https://github.com/prefecthq/marvin ), Vercel AI SDK, and similar efforts.
Right now this amounts to me tinkering away in Python and trying different approaches, is rewarding all by itself. If I manage to get it into a coherent library that could be useful to others I’ll open source it. One of the things I do not want is try and commercialize, because I find many commercial open source projects serving the business first, users second (but this is a rant for another time).