Biome

Biome is a web -app built to provide workouts personalized with your specific goals

Comment

NOTE: Please watch the video on YouTube instead of the Devpost media player, the quality is significantly worse on Devpost.

Inspiration

Online I see many different types of people working out with different goals in mind, but they all have a different starting point. Some people are lean, but some people start on the heavier side, so Biome generates a workout for everyone.

What it does

Asks the user basic onboarding information, and along with the user provided information, Biome uses a VLM (Vision-Language Model) to analyze a photo of the users body to measure their current fitness level. It uses this analysis and data to construct a complete personalized workout for the user to follow based on their preferences.

How we built it

We used python to run the backend, and during the onboarding process, the data is stored to an anonymous id, and then transferred to the users profile after the login with OAuth. After the user uploads a photo, Qwen3 (a VLM) analyzes the image and generates an extremely thorough description of whatever it sees in the image. It is then sent to Open AI along with the data the user entered in the onboarding process to generate a workout plan personalized for the user, along with other metrics like bodyfat percentage, which stores the progress of the user after each image is uploaded in a database.

Challenges we ran into

Getting the VLM to cooperate with our environments was difficult, especially since each of us had a different setup. It was a new tool for all of us, so actually making the script and testing it was the hardest part. Getting the model to generate an accurate description based on the image was also difficult, how in-depth should we go without causing delays in the workout generation. At the end we were able to put together a balance between efficiency and quality outputs.

Accomplishments that we're proud of

Definitely getting the VLM to do what we intended to do. We went into this hackathon expecting to not get a good result but instead gain experience for future hackathons, or maybe even continue this project further. In the end we were happy with what we made, especially on the technical side of things.

What we learned

We spent way to long tuning the frontend to meet our expectations, to focus on a more technical perspective it would have been better to perfect the VLM model further, to ensure quality consistent outputs. Our communication was also lacking, the work was not split evenly and no one knew what to work on, so even though there were four of us, we didn't use our four brains as well as we should have.

What's next for Biome

Currently, a web-app makes it annoying to upload photos as files, a mobile app where the user can click a "take picture" button and take a picture right then and there would be a definite upgrade to our current website. It makes the software more accessible, so they can view their workouts when they are at the gym, or even view their metrics and progress with simply clicking on an app.

Built With

Share this project:

Updates