Modernizing audience segmentation at lululemon

I’ll keep this high-level - out of respect for my employer, it’s the shape of the problem and how we approached it, not the internals.

At lululemon, our team runs PAVE - an ML platform that scores millions of guest profiles to personalize marketing. And like a lot of real-world ML, the hard part was never the models. It was everything around them. Data scientists were spending their days babysitting deployments instead of building, and good work took too long to reach production.

So the question I cared about wasn’t can we build better models? It was why is it so expensive to ship the ones we already know how to build? Reframing from modeling to delivery is what the whole effort turned on.

What we changed

We rebuilt how work moves from a notebook to production:

• Parameterized pipelines (DABs) - templated jobs anyone on the team can ship, instead of bespoke deployments only their author understands.
• A governed feature store on Unity Catalog - one reusable source of truth for features, so the same signal isn’t re-derived, and re-broken, team by team.
• MLflow 3.0 - one consistent way to track, compare, and promote models. “Which version is live, and why” stops being a mystery.
• Agentic AI, next - moving from predictive segmentation toward instrumented decisioning: agents that reason over model outputs to recommend the next best action, not just the next best score.

The throughline is the one I care about everywhere: make a change cheap to make, hard to break, and boring to ship. The win isn’t a cleaner architecture diagram - it’s data scientists getting their time back, and good ideas reaching guests faster.

My role

I co-led this work and co-presented it at Databricks’ Data + AI Summit 2026. My center of gravity is the decision-intelligence side - turning messy, ambiguous problems into models and tools that help people make better calls - with a soft spot for causal inference: uplift, propensity, incrementality. Not just what happened, but why, and what to do next.

If ML delivery, decision intelligence, or causal inference at scale is your world too, I’d love to connect.