Building Netflix - from movie magic to ML reality
summary
Netflix’s “It’s What’s Inside – The Game” runs real‑time face modelling (LivePortrait) and retrieval‑based voice conversion, puppeting friends’ video/audio at 24 fps by running inference on containerized A100s in EKS and LiveKit streams. The build uses Chalice microservices, Claude‑generated scaffolding (about 40% infra code, 95% room logic), reserves 84 P4d‑class GPUs via ML capacity blocks to hit ~500 concurrent users, trains voice models in ~4–5 minutes, and achieves ~42% cost savings versus on‑demand. It works live at scale.
AI-written summary · about
chapters (12)
- 0:00Welcome and introduction
- 1:34Introducing the project team
- 2:57Explaining the movie concept
- 4:00Demo of the body swap game
- 6:30Ideation and project planning
- 9:00Overcoming real-time challenges
- 12:07Video and voice cloning tech
- 14:53Scaling infrastructure with GPUs
- 18:06Architecture and resource management
- 21:41GPU allocation and cleanup
- 25:12Results and project outcomes
- 27:19Key takeaways and summary
related talks
04mInterview
32mRebuilding the AI Platform: Designing for Today’s Chaos and Tomorrow’s Unknowns
53mKeynote Interview
38mNetflix's Ray Platform: From Deep Learning to GenAI
30mHuman-Centric ML Infrastructure at Netflix
37mAsync Inference at Netflix: Batch and Near-Real-Time Serving
29mFrom encoding to machine learning a story of managing media data at Netflix
41mRecommendations for Building Machine Learning Software