Organic Clusters (OG)

Inspiration

When I say I bookmark everything, I mean everything. If I ever find something shocking, awesome, hilarious, drippy, or what have you, I almost always have the primal urge to bottle that moment and try to save it for later. As a result of these tendencies, however, I have a bottomless store of bookmarks and screenshots that have sat collecting dust for nearly a decade. I always want to go through and relive those moments, but the task of sorting through everything has always been so daunting, which is why I bring you today's product.

What it does

Organic Clustering (OG) brings an automatic bookmark and image organization tool, while having the adaptability of learning from user input simultaneously. It will categorize items as it sees fit, until the user makes a move. Each change the user makes is directly logged in the model's reward system, and this immediate learning allows it to instantly adapt and start clustering items in a way the user prefers.

How we built it

At its core, it is a contextual bandit algorithm that has each existing cluster (and the option of a new one) as its arms. When a user changes some cluster/category assignment, the model is penalized and learns this change. It also comes with preexisting configurations to make intelligent similarity-based clustering right out of the box.

Each image and webpage is embedded through the CLIP embedding model, and this is served to the user through an interactive react flow UI.

Challenges we ran into

Many, namely deployment related issues. I'll certainly tackle these after the hackathon though!

Built With

• cloudflare
• python
• react
• vercel