Production MLOps for Spam Detection: Real-Time Multi-Modal AI
summary
System operates at billion-user scale making millions of evaluations per second and targets ~80 millisecond end-to-end latency while keeping a 0.1% false positive rate from affecting millions. Pipeline uses lightweight prefilters, specialist models (text/image/video/behavior), multimodal fusion with cross-attention and a metalerner; feature stack includes transformer embeddings, CNNs, OCR and behavioral signals. Infrastructure uses GPUs/TPUs, quantization, batching, a low-latency feature store, shadow deployments and progressive rollouts; privacy via federated learning and homomorphic encryption, and coordinated attacks detected with graph neural networks.
AI-written summary · about
chapters (13)
- 0:00Introduction to Massive Scale Spam Detection
- 1:58Understanding the Challenges of Billion-User Scale
- 3:27End-to-End ML Ops Architecture
- 4:46Multi-Modal Feature Engineering
- 5:57Model Architecture Ensemble Approach
- 7:14Distributed Serving Architecture
- 8:23Smart Sampling and Learning Strategies
- 9:25Coordinated Attack Detection
- 10:15Ensuring Privacy and Compliance
- 11:10Continuous Model Training and Monitoring
- 12:52Anomaly Detection and Incident Response
- 13:37Future Challenges in Spam Detection
- 14:33Key Takeaways and Conclusion
related talks
19mMLOps End-to-End Speedrun: Build an APIs around a Spam Classifier Model
53mExperience with an end-to-end MLOps pipeline: Practical insights and 3D architectural perspectives
25mReal-Time Fraud Detection at Scale
20mCan Research Drive Product Value? Lessons From Spam Detection
42mOnline Safety Bill: How Global Platforms Use MLOps to Keep People Safe
18mDesigning a scalable ML Ops Pipeline: Insights and best practices
15mIP Detection at Scale: CI/CD Meets Cloud-Native Enforcement
24mMachine Intelligence gets real