From Shadows to Spotlight - How Swiss Post Performs Reliable ML Deployment
summary
Swiss Post moves its Iris fraud service from research to production with a reproducible pipeline: data rooms on S3, SageMaker pipelines for training in a segregated training account, cross‑account model copy to an inference account, and a Spring Boot backend on EKS behind API management. New models run as a shadow model deployment that logs predictions to S3; synthetic labeled requests plus live traffic measure prediction, score and performance drift. A Lambda computes metrics and triggers automatic promotion via Lambda when the shadow beats production; the system goes live in 2025.
AI-written summary · about
chapters (12)
- 0:00Reliable ML deployment intro
- 6:52Iris project evolution
- 8:59Automating ML with SageMaker
- 10:45Data governance & secure rooms
- 13:10AWS ML architecture
- 16:38Continuous model updates
- 18:22Shadow model deployment
- 23:16Model drift in production
- 27:11Validating with synthetic data
- 29:19Promoting shadow models
- 31:04Key takeaways
- 32:48Q&A on training & fraud
related talks
34mFrom Shadows to Spotlight - How Swiss Post Performs Reliable ML Deployment
37mThe story of IRES - How Swiss Post fights fraud with ML on AWS
35mHow Swisscom efficiently fine-tunes and deploys LLMs with Amazon SageMaker
36mInformation Reliability Service - a novel approach to fraud detection
43mMachine Learning Ops: DevOps for Data Science
32mMaking Happy Modelers: Build and Maintain...
55mBuilding a secure MLOps pipeline, featuring PathAI
34mLat Designing and Building Complex Machine Learning....