Been meaning to build similar. Specifically with a view towards mixing local LLMs and openAI. Locals are cheaper and more interesting but less powerful. eg some tasks will need gbnf grammar some won’t. Some can get away with 3.5 others need 4
Suspect I’ll end up duplicating OPs efforts though because hosted doesn’t really work for that.
Sounds interesting - feel free to reach out or discuss on the Braintrust discord :)!
The AI Proxy from the post is for using multiple local and remote LLMs over the OpenAI API (along with API key management apparently), while LocalAI is only for using local LLMs.