I had a friend who was paying someone to write SQL queries for their bespoke accounting software, I read a few of the recent papers on text to sql and thought it would be possible to translate natural language to SQL using some of the better more recent open source models. This is a small proof of concept I created which tries to be as privacy preserving as possible, the sqlite database runs in your browser and sends the schema along with the natural language query to a model I have fine tuned on natural language to sql pairs.

It works surprisingly well, for example with the default chinook database you can query "get the 10 customers with the most revenue" and it will translate that to a query which has the formula for revenue and then it executes this query against the db in your browser. It does this without sending any of the database data to the server, only the schema.