Automated product enrichment pipeline
A beauty retailer's catalog operation went from a three-day manual refresh to a same-day automated pipeline — with more consistent listings, not fewer.
Representative engagement — the figures below model a typical project of this kind, not a named-client audited result.
A catalog growing faster than the team could write it
Product descriptions, structured attributes, and SEO metadata were written by hand. A full catalog refresh took roughly three days of the team's effort, and new supplier drops sat in a queue for five to ten days before going live.
Quality drifted, too. Voice varied between writers, and around 22% of listings were missing or had inconsistent structured attributes — which quietly hurt on-site filtering and search.
A pipeline that enriches, checks itself, and asks for help
We built an enrichment pipeline that ingests raw supplier feeds, normalizes attributes against the retailer's own taxonomy, and generates on-brand descriptions, structured attributes, and SEO metadata.
Every item passes an automated quality and tone evaluation. Items that score below the confidence threshold — about 8% — are routed to a human review queue inside the tools the team already used. Nothing publishes unchecked.
- Supplier-feed ingestion (CSV and API) with attribute normalization
- LLM enrichment for copy, attributes, and SEO metadata
- An evaluation harness scoring tone and completeness
- A low-confidence review queue wired into existing tools
What changed, by the numbers.
The team stopped typing and started merchandising
Catalog refreshes that used to consume three days now run in about two hours, with more complete and more consistent listings than the manual process ever produced. The catalog team's time shifted from writing copy to merchandising decisions — the work that actually moves revenue.