The problem
The integration catalog presented the same recommendations to every user. High-value integrations were buried. No signal fed back from user behavior into what got surfaced.
What I did
Coached an analyst through the full lifecycle: problem framing, behavioral modeling, recommendation engine design, and production deployment. Built in Python. Shipped end-to-end without a dedicated engineering team. The coaching approach was deliberate — the analyst owned the model, not just the analysis.
Outcome
1analyst coached to production ML
↑marketplace conversion rate
0dedicated eng team
The marketplace could surface the right integrations at the right moment in each user's lifecycle, and the analyst had a repeatable ML delivery capability that outlasted the project.