Show HN: I indexed 37h of my videos using an RTX 4090 and local ML models in 24h
A developer indexed 37 hours of video content including podcasts, coding tutorials, and screen recordings using a self-hosted Docker setup with an NVIDIA RTX 4090, achieving significantly faster processing times than on an M1 Max. The longest video—a 3-hour 12-minute livestream—was indexed in 1 hour 52 minutes, analyzing 4,612 frames.
Background
- The author built a local video indexing pipeline that uses an RTX 4090 GPU to run ML models (likely vision, OCR, audio transcription) on personal video archives — podcasts, coding tutorials, screen recordings, GoPro footage.
- This is a follow-up to an earlier post where they ran the same app on an Apple M1 Max chip; the new version uses Docker + NVIDIA and is significantly faster. They processed 37 hours of video in 24 hours.
- The post is notable in the self-hosted / local-AI community because it demonstrates that consumer-grade hardware (a ~$1600 GPU) can handle real-world, messy video at better-than-realtime speed without cloud APIs.
- Key context: most video indexing services (Google Photos, Apple Photos, AWS Rekognition) run in the cloud. Doing this locally with open-source models is technically difficult and appeals to privacy-conscious users and AI tinkerers.