Except they're not.

What they're building is a wrapper around OpenAI's API with an Azerbaijani interface. A Jupyter notebook that runs on someone's laptop. A demo that breaks the moment GPT changes its pricing.

I know because I tried to build something real.

When I started working on NLP, I expected to find tools. Tokenizers, morphological analyzers, spell checkers. Something production-ready I could build on.

There was nothing. Scattered GitHub repos, unmaintained libraries, academic papers with no code. A language spoken by 10 million people and zero unified infrastructure.

So I had to start build it myself.

That's when I understood the real problem. It's not that we lack smart people. It's that everyone skipped the foundation and jumped straight to the flashy part.

The foundation nobody is building: - Language-specific NLP tools that actually work in production - Local models for sensitive data — because banks cannot send their documents to OpenAI - Real infrastructure that doesn't break when a US company changes its API

A Jupyter notebook is not a product. An OpenAI wrapper is not AI. It's renting intelligence from abroad and calling it your own.

Real AI sovereignty starts one level deeper with the infrastructure layer that everyone ignores because it's hard, slow, and unglamorous.

If you've done something similar, I'd love to hear how you approached it.

Star the repo if you find it useful -> github.com/BarsNLP/barsnlp

Also if you work with Azerbaijani text or any low-resource Turkic language, I'd love to hear what you're building. Let's talk.