The story of IRES - How Swiss Post fights fraud with ML on AWS
summary
Swiss Post replaces mailed verification with a real‑time ML trust score named IRES trust score, built on AWS Glue, Athena and Amazon SageMaker and deployed as SageMaker endpoints behind a Kubernetes backend. Training uses a random forest on labeled expert-reviewed data (≈140,000 test rows with ≈1,400 frauds) reaching ~99% recall. Pipeline automates ingestion, retraining, shadow‑variant testing and rollout, goes live 2025, saves 9,000 letters, cuts abandon rate 75% and lifts completions 22%.
AI-written summary · about
chapters (16)
- 0:00Introduction to fraud fight
- 2:11Swiss Post’s fraud challenges
- 3:13Legacy verification flow
- 4:55Redesigned activation flow
- 5:59Data integration in AWS
- 9:09Enterprise data sharing
- 10:56Data preparation pipeline
- 13:01Choosing the ML model
- 14:27Amazon SageMaker overview
- 15:53Training job workflow
- 20:06Model performance metrics
- 22:12Deploying model to production
- 27:00Why automation matters
- 28:03Shadow variant testing
- 33:34Business impact and results
- 35:20Key takeaways summary
related talks
36mInformation Reliability Service - a novel approach to fraud detection
43mFrom Shadows to Spotlight - How Swiss Post Performs Reliable ML Deployment
34mFrom Shadows to Spotlight - How Swiss Post Performs Reliable ML Deployment
27mDetecting fraud faster with Amazon SageMaker, featuring Coinbase
40mFighting Financial Crimes as an SRE
39mFraud Detection: Using ML to Identify and Manage Fraudulent Activities
52mPreparing data for reporting, ML, and AI on AWS
24mAccelerate your business with AWS AI Services