The media processing pipelines behind AI
summary
The team integrates AI inference into existing media pipelines so uploads hit ABR encodings + CDN while also letting the system add a hat to all the dogs in the scene; the platform ingests 1 BILLION media assets DAYLY and serves 1TRILLION views every 24h. Implementation uses ffmpeg one-liner decoding, Python queues, Detectron for segmentation, GPU-backed models with preload and streaming plus backpressure. Surprises: model loads take seconds–minutes for 30s videos, CPU sits idle during GPU inference, and sometimes CPU decode outperforms GPU. A Colab prototype works; production needs custom tradeoffs.
AI-written summary · about
chapters (7)
- 0:00Introduction and agenda
- 1:23AI in media products
- 4:14Traditional media platforms
- 6:36Designing AI pipelines
- 11:06Optimizing pipeline efficiency
- 15:48Hands-on implementation
- 20:14Conclusions and trade-offs
