SkipVid: AI-Powered YouTube Navigation

The Problem

YouTube videos are full of filler. Intros, sponsor segments, recaps, and off-topic tangents eat up huge chunks of your time. You know the answer to your question is somewhere in that 45-minute video, but you're stuck scrubbing through the timeline trying to find it.

I built SkipVid to fix this. It's a Chrome extension backed by a FastAPI server that uses a combination of machine learning, natural language processing, and large language models to give you intelligent control over any YouTube video.

What It Does

Auto-skip intros and sponsors - An XGBoost model trained on video transcript patterns detects intro/sponsor segments and skips them automatically
Keyword search - Jump to exact moments where specific words are spoken (free tier)
Concept search - Find topics even when the exact words aren't used, powered by spaCy NLP
Search-AI - Ask natural language questions like "where does he talk about the database setup?" and get timestamps
Ask-AI - Have a full conversation with the video content using an LLM

Architecture

Backend (FastAPI + Python)

The backend runs on Python 3.11 with FastAPI, deployed on Render via Docker. It handles all the heavy lifting: transcript fetching, ML inference, NLP processing, and LLM orchestration.

# Key backend components
├── app/
│   ├── main.py              # FastAPI app, routes, middleware
│   ├── routes_skip.py       # /api/skip - ML + SponsorBlock fusion
│   ├── routes_semantic_free.py  # Concept search (spaCy)
│   ├── routes_stripe.py     # Stripe checkout + webhooks
│   ├── deps.py              # Auth dependencies
│   └── security.py          # JWT handling
├── intro_detector/
│   └── model/               # XGBoost trained model (92.9% accuracy)
└── Dockerfile

Chrome Extension (MV3)

The extension uses Manifest V3 with a shadow DOM UI injected into YouTube pages. This keeps the extension's styles completely isolated from YouTube's CSS. A service worker (bg.js) handles all API communication, and the content script manages the UI and video player integration.

The Intro Detection Model

One of the most interesting parts of the project. I trained an XGBoost classifier on transcript patterns to detect intro and sponsor segments. The model looks at features like:

Position within the video (intros tend to be in the first 10-15%)
Transcript text patterns (sponsor language, subscribe/like prompts)
Segment duration and timing characteristics
SponsorBlock community data for training labels

The result: 92.9% accuracy on the test set. The model is fused with live SponsorBlock API data at inference time for the best of both worlds - ML predictions for videos without community labels, and verified community data when available.

Auth & Monetization

Authentication flows through Supabase (Google OAuth) with JWT tokens. The extension stores license tokens locally and validates them against the backend. Payments are handled through Stripe with a freemium model:

Tier Structure

Free: Keyword search + auto-skip
Regular ($5/mo): + Concept search + Search-AI
Pro ($10/mo): + Ask-AI chat
All paid tiers include a 30-day free trial.

Challenges & Lessons

Building a Chrome extension that integrates deeply with YouTube taught me a lot about working with third-party platforms:

YouTube's SPA navigation - YouTube doesn't do full page loads, so the extension needs to detect navigation via the History API and MutationObserver
Chrome Web Store review process - Updates can take days to get approved, which creates awkward windows where the extension and backend are out of sync
Shadow DOM isolation - Essential for keeping YouTube's aggressive CSS from breaking the extension UI, but it means you can't use global stylesheets
Rate limiting LLM calls - DeepSeek API costs add up fast with chatty users, so careful prompt engineering and caching are critical

Stack Summary

Backend: Python 3.11, FastAPI, Docker, Render
ML: XGBoost, scikit-learn, spaCy
LLM: DeepSeek API for Search-AI and Ask-AI
Auth: Supabase (Google OAuth), JWT, license tokens
Payments: Stripe subscriptions with webhook handling
Extension: Chrome MV3, shadow DOM, service worker