AntiRot

Inspiration

YouTube is one of the most valuable learning platforms on the internet, and also one of the most addictive. Existing blockers take a blunt approach: block everything. I wanted something smarter.

What it does

AntiRot uses AI to classify YouTube videos in real-time as valuable or time-wasting. Watch a lecture on thermodynamics? Fine. Fall into a 2-hour brain-rot spiral? Blocked. It's a productivity layer that understands context, not just URLs.

How we built it

A Chrome extension captures the video URL and sends it to a FastAPI backend server. The backend calls Supadata's API to extract the video transcript, then passes it to Amazon Nova via OpenRouter. Nova classifies the video and returns a binary signal, which the extension uses to either allow the video or replace the page with a blocking screen.

Challenges we ran into

In v0.1, we used yt-dlp for transcript extraction, it worked perfectly locally. When we deployed to a cloud server in v0.2, YouTube's bot detection blocked all requests from data center IPs. We had to rearchitect the transcript pipeline entirely, switching to Supadata's API.

Accomplishments that we're proud of

Built a full-stack AI product, Chrome extension to cloud backendm solo, in under 48 hours. The system handles YouTube's SPA architecture gracefully and degrades cleanly when APIs are unavailable so it never breaks the core YouTube experience.

What we learned

Building and deploying production APIs with FastAPI + Docker on GCP
Navigating platform bot detection and designing around it
Prompt engineering for reliable binary classification with Nova 2 Lite
Building Chrome extensions with Manifest V3

What's next for AntiRot

Expanding to Instagram, TikTok, and X using Supadata's multi-platform transcript support. Introducing user-defined preferences so the AI learns your definition of valuable. Adding a paid tier with mobile and cross-device sync.