On June 16 I had the privilege of standing on stage at Databricks’ Data + AI Summit 2026 in San Francisco, co-presenting some of the work my team has been doing at lululemon.
The short version: our audience data science team runs PAVE, an ML platform that scores millions of guest profiles to personalize marketing. And like a lot of real-world ML, the most time-consuming part was never really the models - it was everything around them. Data scientists were spending their days babysitting deployments instead of building, and good work took far too long to reach production.
So the talk was about how we changed that:
- parameterized pipelines (DABs) that anyone on the team can ship,
- an enterprise guest feature store on Unity Catalog,
- MLflow 3.0 to unify how we track and promote models,
- and where we think agentic AI takes all of this next.
A few takeaways
Despite being the last talk of the evening, the talk went well, and the conversations afterward were the best part. A few things stuck with me:
- The demand for a simple, API-driven feature store is still unmet in 2026. A point-in-time-correct framework that balances flexibility, performance, and ease of use clearly resonated - though the honest counterpoint I kept hearing is that shared preprocessing across teams is still rare in practice. The
guest_feature_storeAPI demo landed especially well, which was a genuinely nice feeling. - People love their metadata. The registry table - documenting each feature store domain (e.g. the email feature store maps to the email domain) with feature names, logic, business descriptions, tags, aggregation types, and incremental refresh strategy - drew real praise. It was inspired by Tecton’s approach of capturing rich column-level documentation in dedicated tables, and it’s a pattern I’d recommend to anyone building a store others have to trust.
On a personal note
Candidly, most of my takeaways were personal. Being part of such a strong speaker lineup was deeply humbling, and I’m grateful for the chance. A big thank you to the team at Databricks - Tri Chiem and Reynolds Pravindev - and to my manager and co-presenter Jeroen Ruissen, who made it a joy to share the stage. I hope to return to DAIS 2027 as a speaker once again.
If any of this is relevant to what you’re building, I’d love to connect.
