Scaling manual tagging accuracy across a 10-person research team
10 researchers · 9 creator sites · 3-round rollout
Publishing / Creator economy · Operations, data quality and process design · Project · via Hugo
An automated book-tagging pipeline failed, so every book in every creator post across nine sites had to be identified and tagged by hand.
With 10 people in one shared tracker, the real risks were duplicated work, inconsistent edition choices and small errors adding up unseen.
Standardised statuses, built a triage pass for orphaned rows, a closed-loop exception process, weekly QA sampling and written edition-selection rules.
99.9% accuracy across thousands of posts, all nine sites kept current, and a QA process that scaled to a later launch.
Documents kept private because they contain client details. Happy to walk through them in an interview.
The QA and exception process scaled to a later site launch without extra headcount.
Overview. The platform helps book creators share their content and earn from their audiences. When automated tagging could not be trusted, I led a 10-person research team to do the work manually, with direct ownership of accuracy.
- Owned end-to-end accuracy for a tagging workflow spanning nine creator sites and thousands of posts.
- Directed 10 researchers identifying the correct book edition from each post.
- Ran a phased rollout: newest posts across every site first, then 50 posts back per site, then a full year back.
Every book had to be matched to a purchasable edition without duplicate entries and without pausing any creator's feed.
What I built
- 01A controlled status vocabulary (Started, In progress, Done, Skipped) enforced with data validation, so no two people tagged the same post.
- 02A recurring triage pass that caught rows with no owner and no status before they became backlog.
- 03A closed-loop exception process for broken covers, unavailable posts, unreleased editions and import failures.
- 04A defined QA process: a fixed weekly sample per researcher, an error taxonomy and a feedback loop into training.
- 05Velocity tracking by round and by site, so lagging sites got more researchers early.
- 06Written edition rules (paperback first, then hardcover; skip if no eligible format) so 10 people made the same call.
- Tagging accuracy99.9%
- Sites kept current9 of 9
- Researchers10
- Replaced ad-hoc spot checks with an auditable QA trail behind the 99.9% figure.
- Delivered current feeds and bookshelves for every site without pausing new content.
More case studies
Full index- 01US digital pharmacy clientDiagnosing the root causes of customer and compliance failures, separating agent performance from systemic issues, and building the governance to improve both safely at scale93% CSAT vs 90% target
- 02EdTech learning marketplaceReducing recurring customer dissatisfaction through frontline capability84% drop in issue-recognition DSATs
- 03EdTech marketplace and premium tile manufacturerCurriculum design, delivery, assessment and post-training feedbackNew agents at 96 to 100% CSAT in their first quarter