Looks nice, very nice.

Why don't you use your local open source llm, without the interaction of big models? I mean, more work, but you don't need to pay your cut to them. Just asking.