The challenge
Product-catalogue matching ran as a month-long batch process for a national retail operation, slow, opaque and impossible to audit at scale. Local-language matching across source, competitor and cross-entity catalogues lacked governance, guardrails and quality measurement.
What we did
Rebuilt matching as a governed generative AI workflow: centralised multi-model access through an AI Gateway, combined Vector Search with evaluation workflows and reranking, embedded Unity Catalog lineage, guardrails, inference tables, checkpointing and auto-resume from the first sprint.
The outcome
Runtime collapsed from roughly 30 days to roughly 3, a 90 percent cycle-time reduction, with every run now governed, measurable and auditable end to end. Unlocked product substitution, assortment optimisation and competitive intelligence at catalogue scale.
A month-long batch process became a governed, three-day generative AI workflow, faster and auditable end to end.
Architecture paired semantic retrieval with evaluation-driven reranking under Gateway controls. Unity Catalog provided fine-grained permissions and lineage; inference tables and MLflow evaluation made match quality and cost observable for human-in-the-loop review.