Problem
McAfee's applied science team builds propensity and pricing models for personalized offers, churn prevention and trial-to-paid conversion. Feature engineering takes 60-80% of project time. Each new model request needs weeks of hand-written SQL over 780+ columns, and schema drift forces weeks of re-testing.
Example from the article: about 10 feature categories took weeks of exploration and sometimes months of deployment, and produced about 1.5K features.
Solution in brief
An agentic pipeline replaces the manual loop. A discovery agent maps feature categories to tables, one specialist agent per table writes and runs feature SQL in parallel, and a synthesis agent builds cross-table interactions. A critic checks leakage, then sample training and SHAP ranking keep only proven top-N features, and only those are materialized on full data. The speaker claims a 60% cut in time and effort.
Questions this case answers
- How to automate feature engineering across many tables with LLM agents for propensity and churn models, with a leakage check and pruning of generated features
- How to find behavioral features in hundreds of columns spread over raw and cleaned tables when the schema changes over time
- How to automate feature engineering for propensity and churn models with LLM agents
- How to cut the weeks of manual SQL needed to build features for a new model request
Tools and methods
- Discovery agent — hypothesis-driven scouting of tables with four Unity Catalog function tools; outputs a routing brief.
- Specialist agents (one table each) — propose features and write SQL for one source table, in parallel.
- Synthesis agent and critic — build cross-table interaction terms; check leakage by EXPLAIN-only validation.
- Sample training, SHAP ranking — train on a 10-20% sample, log to MLflow, keep top-N features.
- OpenAI Agents SDK, Databricks Foundation Models (Claude Sonnet 4.6), MLflow, Databricks Apps — orchestration, inference, tracing and runtime.
Google or a one-time email code. No password.