CovenantSentinel

CovenantSentinel is an agentic covenant-compliance auditor for credit teams. Banks and private-credit funds check loan covenants by hand every quarter — reading a 40-page agreement, finding the right figures (LTM vs quarterly!), checking whether a preliminary number was superseded, computing, comparing — across hundreds of borrowers. It's slow, error-prone, and a missed breach is a real loss. CovenantSentinel does this like a senior analyst that doubts itself and proves its work. A LangGraph agent plans the audit, retrieves across documents (multi-turn), computes every ratio with deterministic Python tools (the LLM never does arithmetic), decides breach / at-risk / conflict, and loops back to retrieve more when it finds a gap. An adversarial Critic then re-checks every finding — verifying each citation mechanically in code — and overturns false positives. It delivers a cited escalation memo (downloadable as a print-ready PDF) with a grounded "Ask the auditor" Q&A. You watch it reason live. On the ACME case it confirms a real 3.70x leverage breach, eliminates a false liquidity breach by citing the contract's supersession clause, and flags interest coverage drifting toward a Q4-2026 breach — every number traced to a tool call and a verbatim quote, with a confidence score that is a published formula, not a guess. Built entirely during the event on Vultr Serverless Inference (Qwen3.5-397B) + VultronRetriever hybrid retrieval. Trust — deterministic math, code-verified citations, formula-based confidence — is what makes an AI auditor usable in regulated finance. That's the moat, and the wedge into a multi-trillion-dollar private-credit market. Live demo (deployed end-to-end on a Vultr Cloud Compute VM): http://95.179.216.167/