LenderBox

LenderBox — Document-to-metric CRE loan-book corpus with a paired natural-language query layer: covenant terms, LTV, DSCR, debt yield, maturity, concentration and risk tier automatically parsed from underlying loan documents and servicing data, each metric carrying a traceable citation back to its source document; plus logged plain-English-to-portfolio-query pairs (e.g. a filter request over property type, maturity year and DSCR threshold, or exposure by metropolitan area) with their verified answers and full deal context, and credit-policy concentration limits expressed as machine-enforceabl…

LenderBox offers (Alternative, Fundamental, Reference) — Whole-borrower-book scale rather than vendor-catalogue scale: each deployment unifies an institution's entire CRE loan book, so corpus volume is the aggregate of client books — est. tens of thousands to hundreds of thousands of loan records across clients, each with parsed covenant and metric fields plus supporting documents. Query logs accrue continuously per deployment..

Training and evaluating text-to-query / function-calling agents over financial ledgers using the logged question-to-verified-answer pairs as ground truth; document-extraction models for loan agreements and covenant structures with span-level supervision; grounded-generation and citation-faithfulness evaluation (every answer must be traceable to source); CRE credit-deterioration and refinance-default features from maturity-wall and DSCR trajectories; regulatory-guideline compliance classification; hallucination benchmarks built on the metric-to-document alignment

The data is with 2 years of history.

Coverage spans US; Financial Exchanges & Data, Commercial Banks; alternative, fundamental, reference; loans, fixed_income, structured_products.