FinIntel AI: A Confirmation-Gated, Explainable Platform for Corporate Financial Analysis, Credit Risk Assessment and BankRouted Lending

Abstract

Small and medium-sized enterprises seldom employ a financial analyst, and the lenders who assess them still work from documents rather than from verified figures. Statements arrive as spreadsheets, portable documents and scanned images; values are re-keyed by hand into private spreadsheets; interim statements are compared against thresholds calibrated on annual accounts; and the credit opinion that results carries neither a recorded probability nor a stated reason. This paper presents FinIntel AI, a role-based web platform that converts uploaded financial statements into confirmed, machine-readable data and derives ratio analysis, financial-health and credit-risk scoring, feature-level explanation, forecasting, simulation, anomaly detection, a grounded question-answering assistant and a bank-routed lending workflow from that single verified dataset. The design rests on a confirmation gate: no analytical module may read a statement whose extracted figures the subject company has not explicitly approved, and every stored result carries the version of the data from which it was computed, so an assessment produced before a correction is surfaced as stale rather than presented as current. A reporting-period model annualises interim statements before any comparison or scoring, removing a distortion that otherwise penalises companies reporting more frequently than annually. Financial health and corporate credit risk are modelled with gradient-boosted decision trees; every prediction is accompanied by an additive feature attribution rendered inside the application; irregular transactions are identified by an unsupervised detector supported by explicit rules; and the assistant answers only from results the platform has already computed, so it cannot invent a figure. A single application reaches one, several or all eligible banks, and each bank decides independently against the assessment snapshot taken at submission, with the aggregate status recomputed from those decisions. The platform was evaluated on held-out partitions of its training data and then exercised end to end on demonstration companies drawn from different industries, with every analytical output and every service response time recorded. Because the classifiers were trained on synthetic data, the reported measurements characterise model behaviour rather than real-world credit accuracy.