Hey that's very cool and works surprisingly well!
At Veezoo (http://www.veezoo.com) we have been tackling this problem for over 5 years now.
Under the hood we are using our own models. With GPT-3 we're a bit worried about the lack of fine-grained control needed for productive use-cases and obv. also lock in.
Will try out against the same dataset and see how it compares!