How We Built Ad Moderation from Scratch and Reached 1 Billion Verdicts per Day
summary
Describes building an ad moderation system at Yandex to process over 1 billion verdicts per day with automated checks on the order of single-digit seconds. Covers system architecture and engineering choices to handle high throughput and low latency while maintaining high moderation quality.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 3:13Moderation systems evolution
- 6:08BigM architecture
- 11:23Combining ad parts
- 15:45Processing atoms and parallel graphs
- 18:15Data testing and labeling
- 20:49Aggregation and decision caching
- 23:27Error handling and resilience
- 25:14Outcomes and scalability
- 25:56Q&A
related talks
33mMachine Learning to moderate ads
39mBuilding a Real-World Adverse Media Pipeline
38mYandex’s Crowdsourcing Ecosystem
16mLeveraging ML to Create a Reliable Anti-Fraud System
33mAdTech
52mHit 350 Milliseconds: The Evolution of Mail.ru Mail Anti-Spam Architecture
30mML Automoderation System for LLM-Augmented Online Platforms
29mAI and ML for Computational Advertising