WatchWise

Scores how overstimulating/brain-rotty YouTube content is for kids, grounded in neuroscience.

Prezi link: https://prezi.com/p/edit/rsyd2wcrpvna/

Devpost link: https://devpost.com/software/rotcheck

Prerequisites

• Python 3.11+
• Node.js 18+
• yt-dlp installed and on PATH
• ffmpeg installed and on PATH
• Anthropic API key

Backend Setup

cd backend
python -m venv venv
venv\Scripts\activate        # Windows
pip install -r requirements.txt

# Copy and fill in your Anthropic key
cp .env.example .env

uvicorn main:app --reload

API runs at http://localhost:8000

Frontend Setup

cd frontend
npm install
npm start

App runs at https://watchwise-blond.vercel.app/

Features

Page	What it does
Score Video	Paste a YouTube URL → BrainRot score (0–100) + radar chart, with live agent streaming
Watch History	Upload Google Takeout watch-history.json → dashboard with trend line + creator table
Creator Lookup	Enter a channel URL → samples 5 recent videos → average kid-friendliness rating

Scoring Dimensions (Radar Chart)

• Pacing — cuts/min from ffmpeg scene detection (falls back to librosa onset detection)
• Sensory Overload — volume spike frequency + avg volume variance via IQR analysis
• Educational Value — transcript analysis via Claude
• Manipulation — clickbait, parasocial language, fake urgency detection
• Dopamine Cycling — variable reward schedule modelling from pacing + audio signals

Age Adjustment

All scoring agents receive the child's age. Thresholds shift — 3 cuts/min is flagged as high pacing for a 4-year-old but acceptable for a 12-year-old.

Architecture

yt-dlp → video (ffmpeg: scene cuts → cuts_per_min)
       → audio (librosa: avg_volume_variance, volume_spike_frequency)
       → transcript (auto-subs) + metadata (title, channel, thumbnail, duration_sec)
            ↓
   pacing_agent          — attentional demand (Fernando FSM model)
   sensory_agent         — audio environment vs clinical thresholds
   educational_agent     — content quality + language complexity
   manipulation_agent    — clickbait, parasocial language, fake urgency
   dopamine_agent        — variable reward schedule risk
            ↓
   judge_agent — synthesizes → BrainRot score (0–100) + verdict + radar + summary

Results stream to the frontend in real-time as each specialist completes. Pipeline output is cached in-memory so repeated URLs skip re-downloading.