How Vivix Scales Video Ad Classification with Ray
summary
Company scales ad classification from 30k to a 6 million label target by piping video into Event Hubs, Spark Structured Streaming on Databricks and Ray for distributed inference, using Whisper for transcription, OCR, vector DB similarity and LLMs as a preprocessing label generator. Through Ray on Databricks and partial-GPU actors throughput rises from ~800 to 2–3k creatives/hour, LLM augmentation yields big accuracy gains (reported LLama2 15% boost and layered models up to ~50% improvement).
AI-written summary · about
chapters (7)
- 0:00Introduction to Vivix
- 4:06Company and problem overview
- 6:41Why Ray for scaling
- 13:14Ray performance improvements
- 13:56Gen AI for classification
- 19:58Data-first AI principle
- 20:39Q&A session
related talks
28mScaling Multimodal Data and Reinforcement Learning with Ray
30mScaling Machine Learning at Discord: Our Journey with Ray and Anyscale
21mHow Ray and Anyscale Make it Easy to do Massive-scale ML on Aerial Imagery
26mPetabyte-Scale Video Curation on CoreWeave
28mServing LLMs for Agents at Scale with Ray + vLLM
44mPanel: ML Platform on Ray
17mScaling and Unifying SciKit Learn and Spark Pipelines using Ray
20mThe Quick Journey to Using Ray: How We Implement Ray and Anyscale to Speed up our ML Processes