Sport Sage

What it does

Sport Sage builds on the powers of Gemini to process short athletic videos. Prompting the user for brief pieces of context we then build a useful response to improve athletic performance. The application targets non-experts and is meant to act as a "rubber duck" level of help in your pocket. We prioritized fast onboarding and minimal feedback. Many folks still find AI and LLMs intimidating, so we strove to hide as much complexity as possible behind an obvious interface. Narrowing the set of things you're asking Gemini for helps contextualize the feedback and makes it nearly instantaneously useful.

The processing and understanding of the video is fairly good in cases where the video is clear and single subject. Team sports are a little more challenging and lead to less specific feedback, but it can still be useful.

After experimenting with the API, we found the default output on one and few-shot requests to be pretty good! The team is experimenting with some additional training data, but this didn't seem necessary for the prototype we built. Adding curated model intelligence are and streaming feedback sessions our primary next steps.

In terms of technologies, we used Elixir and Phoenix (with LiveView) to develop the web application.