Case study — Metabooks
Making ebooks findable.
An AI-native platform that turns raw book metadata into platform-ready, discoverable listings — across the channels readers actually use to find them.
The problem.
Ebook distribution depends on metadata — titles, descriptions, keywords, category codes — and most of what publishers send is incomplete, inconsistently formatted, and missing the platform-specific classifications that determine whether a book shows up when a reader is looking for it.
The stakes are real. Poor metadata means poor discoverability, which means poor sales — and for a distributor, that reflects on every publisher relationship they hold. Doing this well by hand across catalogs of thousands of titles is not realistic. So in practice it doesn't get done, and books that should be findable, aren't.
What we built.
A SaaS platform that takes a publisher's catalog, enriches each title with the context the AI needs, and generates optimised metadata across four dimensions: descriptions written for how each platform actually renders them, subtitles that carry search intent, category codes that map onto the taxonomies distribution platforms rank by, and SEO keywords grounded in real Amazon marketplace data — bid signals pulled directly from Amazon's ads API rather than guessed at by the model.
A reviewer accepts, rejects, or refines each suggestion — the AI proposes, the editor decides. Every generation is versioned, so no output is ever lost and the current live metadata is always clearly separated from a draft the team is still working on. Approved versions roll into the catalog and export cleanly to the formats distribution platforms expect.
Import handles the several publisher spreadsheet formats already in daily use — the system identifies the format automatically, so publishers don't have to change how they work.
We designed and built it end to end — backend, frontend, AI workflows, infrastructure.
Why it was hard.
The easy version of this product is a chatbot that rewrites blurbs. The useful version is much harder: it has to know something specific about each book before it writes anything, because generic AI output is exactly what publishers already produce by hand — and it isn't helping.
So the platform's real work happens before the AI ever runs. Each title is enriched with context the model can actually reason from, and that context flows into every prompt. The output is specific to the book, not generically plausible.
The second constraint is editorial trust. Publishers don't want their catalog rewritten by a machine and shipped without a human in the loop. So the workflow is built around versioning: every AI suggestion is a draft, sitting alongside the current live data, until an editor decides. Nothing goes live silently, nothing is ever lost, and the audit trail is complete.
The outcome.
A working platform turning a manual, impossible-at-scale task into something a small team can run across thousands of titles — with the AI handling the volume and humans staying in control of editorial quality.
Designed and built solo, end-to-end, in around two and a half months.
Engagement.
Open Build. Designed and delivered end to end — product engineering, AI workflows, infrastructure, deployment.
AI work on real catalogs?
Practical AI on messy, large-scale data is the kind of work we like.
Let's talkLean builders. We deliver.