01 · The challenge
The comps drove real decisions. The logic was a black box.
In publishing, a comp is a comparable title whose past performance is used to plan a new one. Comps drive both marketing and supply chain decisions.
In the existing planning platform, comp selection ran on overnight batches, the selection logic was opaque, business users needed IT to make overrides, and licensing cost was significant.
02 · What I did
What I owned and built.
- Designed a hybrid architecture. SQL hard filters enforce the non-negotiable business rules first (channel routing, key attribute, budget category, merchandising subject), then vector search matches on theme, tone, audience, and other qualitative attributes.
- Built a comp-lookup tool that takes batches of UPC/ISBN codes and returns ranked comparable titles, each with an explanation of the match.
- Encoded channel integrity (a retailer-exclusive title only compares to titles from the same channel) and a rule that a title is never its own comp, which matters for re-releases that share a UPC.
- Presented the system to the marketing team and used their “not another tool” feedback to steer delivery toward automated outputs inside their existing workflow.
- Positioned the work as complementary to the existing platform rather than a replacement, and submitted a 40-minute breakout proposal to a major data & AI conference.
03 · Architecture
How it fits together.
Channel · key attribute · budget category · subject
Theme · tone · audience
→ planner workflow / planning platform
04 · Key decisions
The judgment behind the build.
Why filters before vectors?
Business rules that can’t be broken shouldn’t be left to probability. Hard filters guarantee eligibility, and vector similarity only ranks within the safe set.
Why explain every match?
Planners won’t trust a comp they can’t reason about. A reason on every result builds confidence and makes overrides informed instead of arbitrary.
Why complement instead of replace?
Replacing a core planning platform is a multi-year conversation. Complementing it delivers value now and builds a path to lower licensing cost later.
Why automated outputs instead of a new UI?
Users told me plainly they didn’t want another tool. Meeting them inside their existing workflow is what actually drives adoption.
05 · Impact
What changed because of the work.
- Removes the overnight batch wait, so planners get comps when they need them.
- Transparent selection logic builds trust with business users.
- Business users can see and control comp choices themselves, which opens a path to lower licensing cost.
06 · What’s next
Where it goes from here.
Add a 24-month sales and store-count history to the index with automated refresh, schedule a daily export to blob storage, and decide between an app upload interface and a fully automated feed into the planning platform.
This case study is a sanitized reconstruction. Organization names, proprietary code, data, and internal schemas are intentionally left out.