Sentinel

AI agents can now take real actions — send memos, escalate to executives, move money. But nothing sits between "the agent decided" and "the agent acted": the review happens after the irreversible move. SENTINEL closes that gap. SENTINEL is an autonomous financial oversight agent behind a fail-closed gate. Given a covenant-monitoring brief, it plans a multi-step investigation, retrieves evidence across three passes from a Vultr Serverless Vector Store (reranked by VultronRetriever), computes the Debt/EBITDA ratio with a deterministic tool, conditionally digs into transaction root causes, and drafts a cited escalation memo. Before the irreversible send, three parallel signals evaluate the action — task drift, an adversarial MAARS probe, and citation completeness. If all pass, the memo executes; if any fails, the action freezes and a human operator decides with full context. Every decision is sealed with a SHA-256 hash — edit the record and the seal breaks. The gate's inputs are domain-agnostic (task brief, action, citations), so the same oversight layer fronts SLA escalations, claims letters, or pricing changes — covenant compliance is the first workload. Two demo scenarios (breach → frozen, compliant → executed) prove the gate genuinely decides. Built on Vultr: VM deployment, Serverless Inference (Qwen3.6-27B reasoning), VultronRetriever reranking.