Natural Interaction - Hey CARV (LLM based with "reasoning"?)

Instructor-like spoken interaction with CARV like:

Hey CARV,

• give me feedback about the last run

• how to improve (SPECIFIC_METRIC|MOST_IMPORTANT_METRIC)

• why is my SPECIFIC_METRIC bad/good?

• which drills could I try to improve (SPECIFIC_METRIC|MOST_IMPORTANT_METRIC)

• etc.

Right now, the spoken feedback feels quite a bit static and repetitive. I'd like to have more adaptive feedback, especially if I'm continuously failing to improve a metric/skill.

I'm personally a Google Gemini power user, so I started to feed CARV screenshots into Gemini to get some reasoning into the feedback and results are not that bad. I just could imagine how a special trained LLM could work out, accessing the real data instead of Gemini using my screenshots 🤓