Problem / opportunity
Lending teams wait days for analysts to answer routine portfolio questions. The data is already modeled โ the bottleneck is access. This app puts a conversational agent on a governed lending gold layer so answers come back in seconds using certified metric definitions.
Who it's for
Credit risk and portfolio managers, collections leads, underwriting and credit policy teams, regional and product heads, and lending executives โ business users who need answers but don't write SQL.
Architecture and data flow
Medallion on the lakehouse. Bronze lands core lending, servicing/payments, bureau, and origination data as-is. Silver cleanses and conforms it into dimension tables. Gold serves four curated products: Br 360, Performance Mart, Delinquency & Aging, and Origination & Channel. A Genie space scoped to gold only powers the app; Unity Catalog governs access and row-level security throughout.
What users can ask
Portfolio health (PAR30, default and collection rates, NPL counts), profiles (credit grades, segments, exposure), origination trends (by product, officer, region, vintage), and payment behavior (installment collections, delinquency patterns). Anything outside the lending book is out of scope.
How Genie powers it
Genie is the interface. It resolves questions against the gold schema, generates and runs SQL, and returns results with the query visible. Certified metric definitions are pinned in the space so the same question always resolves the same way, and multi-turn context lets users drill down conversationally.
What we learned
Gold-layer quality matters more than prompt tuning โ narrowing the space to purpose-built marts fixed most wrong answers. Business vocabulary has to be defined explicitly, since "delinquent" and "at risk" mean different things to different teams. Financial metrics must be precomputed, not derived on the fly, so results reconcile with official reporting. And a benchmark question set with known answers is what turns "seems good" into a measurable accuracy number.
Amit Sharma
