SephoraAI Engineer (Contract)

Overview

Sephora's distribution-center leadership was getting inventory answers the slow way: someone pulled a report, someone else interpreted it, and by the time it landed the question had changed. I built a production agent over Databricks that answers questions about inventory health, availability, and risk directly, cutting manual reporting by roughly 70%.

What I built

• A grounded agent. Indexed 150+ inventory SOPs (Blue Yonder) into a hybrid BM25 plus vector RAG store, so every answer and every generated Databricks SQL query was tied to the actual operating procedures and definitions.
• Production deployment. Shipped it on Sephora's Databricks cluster, including fixing JDBC connectivity and timeout failures against Oracle.
• A feedback loop. Executives rated each answer, and I used that to tune retrieval and the SQL-generation prompts over the engagement.

This was a contract alongside Lucile, and it was my first time shipping an agent inside a large company's data stack rather than on infrastructure I controlled.