loan-document-intelligence

loan-document-intelligence is a document-intelligence agent for retail-lending underwriting. It extracts income and bank-statement data from an applicant's documents and runs deterministic cross-validation across them, including declared income versus payslip versus bank-statement salary credits, name/address consistency, balance trend, and affordability. It produces a cited, audited, maker-checker-gated income verification. It is decision-support for underwriting, not a lending decision; the agent verifies and the underwriter decides. It handles applicant PII, so the full R1 redaction and guardrail pipeline applies. The system is built ports-and-adapters on the Gemini Enterprise Agent Platform in the asia-southeast1 region, with an offline test and lint gate that runs without the Google Cloud SDK installed. It produces artifacts including LoanApplicationCase, CrossValidationResult, and IncomeVerificationSummary. Every figure is cited to a source document and field, every interaction is written to a WORM audit log, and the LLM only normalises and explains; it never overrides a deterministic check verdict. Profiles include local, live, gcp, platform, and onprem. The local profile is a real, SDK-free laptop stack with no API key, no emulator, and no google-cloud-* package. The project is licensed Apache-2.0.