John Thomas, IBM Data and AI | IBM Data and AI Forum
summary
Argues that production-grade ML needs software-style devops: CI/CD for models plus asset promotion, monitoring against business KPIs, and automated retraining. Shows a client case with Wunderman Thompson using autoAI selecting and composing features to cut 20,000→2,000 features, then combining transforms to boost signal. Insists on transaction-level explainability and recorded data lineage so each individual score can be replayed, audited and judged against a business-defined fairness bracket for compliance and consumer questions.
AI-written summary · about
