Sovereign Labs

CovenantOps Agent is a web-based enterprise finance agent built for the RAISE Hackathon 2026 Vultr Track. It helps credit teams monitor loan covenants across real financial documents and produce escalation memos that are traceable, governed, and independently verifiable. The system ingests credit agreements, waiver letters, management accounts, transaction data, and supporting evidence in formats such as PDF, DOCX, XLSX, CSV, and OCR-supported images. It then plans the covenant review workflow, retrieves the relevant covenant clauses, extracts financial evidence, recalculates key ratios such as leverage, interest cover, and liquidity, applies valid waivers, and checks transaction-level causes behind flagged movements. Unlike a simple RAG application, CovenantOps Agent performs a multi-step credit-monitoring workflow with deterministic ratio calculations, context-integrity checks, source trust scoring, recovery checkpoints, and governed tool execution. It also generates a signed escalation memo containing the covenant status, ratio versus threshold, headroom, likely causes, confidence level, recommendations, and citations. A key differentiator is verifiability. Each memo is linked to an Ed25519-signed receipt and SHA-256 evidence hash, allowing reviewers to detect tampering and verify the output offline without trusting the application server. The system also includes poisoning-gated self-improvement, meaning it can learn from trusted runs while preventing low-confidence, stale, blocked, or suspicious runs from influencing future outputs. The project demonstrates how enterprise AI agents can move beyond document search into auditable, resumable, and verifiable workflows for real finance operations.