From Catalog Fragmentation to Procurement Visibility: An Agentic Reference Architecture for Puerto Rico's First Government Item Master

Abstract

The General Services Administration (ASG) is Puerto Rico's central procurement agency, yet its catalog contains hundreds of thousands of records with extensive duplication: a single blue ballpoint pen appears in over a hundred textual variations. This paper presents the design and implementation of an agentic Big Data infrastructure that converts the fragmented catalog into a controlled single source of truth. The architecture combines a Databricks Medallion Lakehouse with Microsoft Foundry agents that produce standardized product names and assign National Institute of Governmental Purchasing (NIGP) codes. Delivered artifacts include a daily ingestion task, cleansing and deduplication logic, five Foundry agents with responsible-AI safeguards, a unique item identifier, and a Minimum Viable Data template. The as-deployed first-year cost (~$72,000 USD) is roughly an order of magnitude below earlier hand-curation quotes. The result is a governance-aligned reference architecture for public-sector catalog modernization.