Every U.S. state describes what students should learn in its own language. If you sell educational content nationally, "is this aligned to our standards?" is the first question a district asks — and our honest answer involved months of manual correlation work. So products shipped without alignment, sales conversations stalled, and every request meant another round of manual correlation work. To me it was obviously a graph problem that everyone was handling in spreadsheets.
What we built
A correlation service that treats the problem as a graph. State standards are hierarchical and heavily cross-referenced; content maps to them many-to-many, with room for judgment. We modeled standards and content as nodes in Amazon Neptune, with correlations as edges carrying confidence scores. A Bedrock adjudication stage returns structured correlation judgments with confidence scores; Step Functions orchestrates the batch pipeline across catalogs and standard revisions.
Design decisions
Model it as a graph. The value is in traversal — "show everything aligned to this strand, in this grade band, across these two state revisions." Those queries are painful in a relational model and straightforward in a graph.
The model proposes; editors approve. Correlation quality is a promise we make to schools, so no correlation ships on model confidence alone. High-confidence matches go to fast review; low-confidence matches route to editorial judgment. The human review step is what made the output trustworthy enough to sell.
Run it as a batch. Correlations change when catalogs or standards change, not on every request. Batch orchestration kept costs predictable and let us re-run whole catalogs when a state revised its standards.
Results
Standards alignment for the magazines catalog completed on schedule, and districts started getting answers the same week they asked. The service returned about $135K in sales that had been projected at zero, and avoided roughly $150K more in projected losses.
I architected the pipeline and led the five-person team that shipped it, spending as much time with product and sales on what "aligned" had to mean for a district to buy it as I spent on the traversal queries. If you want the technical warm-up that preceded this, I wrote about exploring Neptune graphs with AI agents.