This is interesting. Funnily enough this is something that we also built in the tool I'm building called nao, which is a data editor for data teams. Because most of the SQL parsers aren't fault tolerant we had to develop our own fault-tolerant parser.

I really think what you showcase in the start of the article is the best thing, being able to statically highlight error in a text form for a user (or for a LLM) to iterate on the fixes is the best thing.

How many AIs do you want in your domain?

Yes