CleanBucket

CleanBucket is an app where people can record themselves recycling in a fun manner, and people can scroll through videos and provide ratings.

Inspiration

The idea for this app came from wanting to make recycling more fun and engaging. Recycling can feel like a chore, and I wanted to find a way to change that. By combining social media elements with sustainability, I thought it would be cool to let people record and share their recycling efforts in a way that's creative and interactive, especially with the rise of short-form content in modern day media. The hope is that by making it social and enjoyable, it’ll motivate more people to get involved and make a real impact.

What it does

The app lets users post videos of themselves recycling, whether they’re showing off a cool shot or just having fun with the process. The rating system would add a bit of a competition, where people can see what others are doing and get inspired by the algorithm. The algorithm utilized multiple layers of AI models and predictive rating in order to show the best videos to the users.

How it all works

Initially, the user opens up an app on their phone and uploads a video of themselves making a cool trickshot into the trash can or recycling bin. Then the video is analyzed using the LLAVA 1.5-7b-hf model hosted on CloudFlare's AI Workspace, and then an initial description is produced. Based on the description, a large number of fetch.ai agents running TinyLLaMa-1.1b-chat-v1.0 analyze the descriptions, have conversations with each other about the descriptions, and write up many comments about the video. Then, a final model is utilized in order to produce an accurate rating based on all of the previous outputs and the videos are then ranked by this rating and sent back to any users who want to view the videos. All of the data communication is done through MongoDB's Atlas except for the videos which are stored on a Cloud server.

Challenges we ran into

Swift is not kind to developers on a limited timeline. The amount of issues we ran into with Swift alone probably matches the amount of bugs and problems we encountered with every other platform combined.

Accomplishments that we're proud of

Getting an AI pipeline working! Thank you to all the help from the mentors, Fetch.ai, and CloudFlare for giving me a taste of what not just one, but many AI agents can accomplish.

What we learned

Pretty much everything was new to us, so it would take too long to detail out everything we learned. But in general, for next time, we should focus on the UI a bit more, think about the data and how it'll move across platforms and devices BEFORE starting to code :).

What's next for CleanShot

We would love to integrate user feedback into the AI loop! Users ranking how cool they thought each videos are would be very useful data, especially for finetuning the final rating system or filtering out mispredictions. We tried to implement this during the hackathon, but unfortunately ran out of time focusing on the AI models. Also, improving the performance of the video playback to reduce delays would also be something to look at.

Built With

• cloudflare
• express.js
• fetch.ai
• llava
• mongodb
• swift