Book metadata enrichment at catalog scale
Workflows covering 30,000+ titles, with model selection, review and distribution constraints.
The problem
Book discovery depends on usable descriptions, keywords and classifications. Updating a large backlist manually creates a substantial editorial workload, while an incorrect update can overwrite information worth keeping.
My responsibility
I developed AI-assisted book-metadata enrichment workflows at Independent Publishers Group. The work covered keywords, BISAC classifications and descriptions, with more than 30,000 titles processed across the workflows.
The approach
The workflow uses book content and existing metadata to propose changes. The engineering decisions include how much content to process, which model to use and how to inspect the difference between an existing record and a proposed update.
Distribution is part of the system: an improved record has to reach the retailer before it can affect discovery. Generated descriptions also need editorial controls so useful publisher copy is preserved.
Result and measurement
The 30,000+ figure describes processing coverage. It is not evidence of an attributable increase in sales. Sales measurement was affected by distribution gaps and incomplete retailer data.
I also developed an internal Docker-hosted PDF catalog generator. This extended automation from metadata processing into assembling a usable sales artifact.
Leadership perspective
The project required trade-offs between processing cost, quality and the ability to review changes. The next business experiment needs a stable delivery path and a comparison group before any revenue effect can be claimed.