Machine Learning Models and Productionization
summary
Few models reach production because real constraints bite: on‑premise insurance deployments require shipping to private clouds with no access, training data is tiny and unrepresentative, and on-premise insurance claims automation forces slow release cycles. Teams build an MLOps stack—MLflow and DVC for model/data versioning, a feature store, logging, A/B routing and rollback—then add daily retrainings and feedback loops; a half‑accurate starter model can cut annotation work ~50%, but refresh cadence still ranges from three days to a month and rollouts require staged traffic and rapid rollback.
AI-written summary · about
chapters (19)
- 0:00Introduction
- 3:02Tell us about your journey
- 6:03Tell us about your challenges
- 9:30What is the difference between ML and DevOps
- 10:12How do you decide if a model is ready for production
- 12:00The opportunity to put one machine learning model into production
- 15:14The challenge of taking risks
- 17:20The feedback loop
- 20:14Deployment
- 24:37Golden Window
- 26:13Refresh
- 27:52Transitions
- 29:06Data Science
- 31:43Model Retraining Downtime
- 38:28Build Graduation
- 46:27Data Versioning
- 47:21Model Washing
- 51:19Alignment
- 53:18Opportunity Costs
related talks
47mMachine Learning in Production: An Intro to MLOps.
31mMLOps In Practice – How To Run Your Machine Learning Models In Production At Enterprise Scale
13mMLOps: From AI Research to Production-Ready Systems
14mGetting Machine Learning applications into production using MLOps and CD4ML
32mAutomating Production Production Level Machine Learning on AWS
13mMLOps at Scale
25mContinuous ML Applications in Production
10mFrom Model to Production: Building Scalable MLOps Pipelines