ATLAS is an agentic AI platform built for financial forensics and regulatory compliance. It orchestrates four specialized agents: a Vision agent that extracts structured data from tax documents via OCR, a Reasoning agent powered by a custom fine-tuned Qwen3-14B (atlas-r3) that detects arithmetic errors, invalid RFCs, and fiscal anomalies, a Validator agent that cross-checks findings against Mexican and US tax law (CFF, LISR, LIVA, FinCEN, CTA), and an Explainer agent that generates human-readable audit reports. The centerpiece is the Regulatory Sandbox — a predictive simulation engine that analyzes a proposed business operation before it executes, surfacing cross-border tax exposure, regulatory timelines, and alternative restructuring scenarios. The model was fine-tuned in two rounds on AMD MI300X, achieving a final eval loss of 0.1016 on a 13,588-example dataset covering Mexican and US fiscal compliance, cross-border transfer pricing, and adversarial red-team scenarios.
Category tags: