Video Scraper
A local pipeline that turns hours-long videos into navigable, timestamped notes — transcription and topic segmentation running entirely on a 4GB GPU, with a private, tunnel-exposed uploader to trigger it from anywhere.
About this project
video-scraper transcribes video locally with faster-whisper and uses a local LLM (via Ollama) to group the transcript into timestamped topic sections, generating a markdown notes file with per-section screenshots and deep links to the exact moment in the video. Built entirely around one hard constraint: a 4GB VRAM GPU. Whisper and the LLM are never loaded together, chunk sizes were tuned by empirical testing instead of maxed out to the model's context window, and every timestamp the LLM returns is validated against the real transcript segments before being trusted — never accepted on faith. A companion project, vscrap, exposes the pipeline to the internet through a password-protected Express server behind a named Cloudflare Tunnel: chunked uploads to get around Cloudflare's 100MB request limit, a hand-checked session guard on the WebSocket upgrade event (which normally bypasses Express middleware entirely), and a backend that only ever binds to 127.0.0.1.