Turning a manual data curation bottleneck into an AI system

The problem

The company collected fine-grained behavioral data (web URLs, app bundle IDs), but their clients wanted to analyze it at a brand level. They needed a hierarchical taxonomy that mapped apps and websites into multilevel company ownership - for example google.com → Google Search → Google → Alphabet, Inc.

The approach

The project started with vague requirements from the sales team. I worked with sales and engineering to define what success actually looked like: an initial set of manually curated brands, a database schema designed for ongoing maintenance, and an AI system to extend the curation at scale.

I started by manually curating a subset of brands with the team, which surfaced the rules and edge cases the work depended on. I then built a library of well-structured database operations to keep the underlying data safe, and translated my manual curation process into Claude skills and agents that could reproduce it across the full catalog.

The outcome

• Unlocked a new commercial offering of brand-level analytics
• Enabled curation of brands at an order of magnitude faster than manual curation, with a smaller price tag
• Durable infrastructure, including a maintainable taxonomy schema and a constrained set of database operations, enabled extending and correcting the system over time