Platform-First Forecasting: Engineering Scalable ML Systems
summary
Claims a platform-first forecasting stack that trades marginal accuracy gains for operational excellence, scalable microservices and business-aligned loss functions. Implements concrete tools and patterns: Kafka for stream ingress, feature stores for reuse, containerized workloads with horizontal pod autoscaling, API versioning and persistent storage policies. Shows deployed systems across e‑commerce, global logistics and finance, cites Amazon-scale production use, reports consistent weekly outputs and reduced business-side failure modes via asymmetric error handling.
AI-written summary · about
chapters (12)
- 0:00Introduction to Platform First Forecasting
- 0:39Speaker Background and Experience
- 1:12Agenda Overview
- 1:48Rethinking Machine Learning Based Forecasting
- 4:52Key Components of Platform-Based Forecasting
- 6:33Success Metrics Beyond Accuracy
- 9:58Asymmetric Error Handling in Forecasting
- 13:03Microservices Architecture for ML Systems
- 15:05Containerization and Orchestration Strategies
- 17:45API Design Best Practices
- 19:20Real-World Implementation Success Stories
- 22:09Future Directions and Conclusion
related talks
21mDon’t Call It 'The Forecast': Structured Prediction at Scale
40mMulti-Series Forecasting at Scale with StatsForecast
14mPerformance Optimization in High-Volume Transaction Systems
26mEngineering Excellence: Platforms and Architectures for Real-Time AI Analytics
29mDemystifying Machine Learning in Production: Reasoning about a Large-Scale ML Platform
31mUse-Case Driven ML Platforms
21mBuilding Scalable ML Systems: Strategies and Insights
33mBuilding Resilient (ML) Pipelines for MLOps