Lucid — AI triage you can actually audit.

About this project

Problem: Fraud and security analysts spend most of their shift on repetitive, manual investigation — pulling device/geo data, correlating account activity, and cross-referencing logs across disconnected tools before they can even make a call. When AI is introduced to speed this up, it usually acts as a black box: a risk score with no visible reasoning. In regulated industries like BFSI and security operations, that's a real liability — decisions need to be explainable and defensible to auditors, regulators, and customers, not just "the model said 87% risk." Solution: Lucid is an MCP-based AI investigation agent, built end-to-end on NitroStack, that runs a full triage pipeline on any flagged alert or transaction: Enrich — pulls device risk, geo-anomaly, and reputation signals for the alert Correlate — surfaces related account activity (logins, resets, prior transactions) Classify — returns a verdict (clear / flag / escalate) with a confidence score and a step-by-step reasoning chain Escalate — auto-drafts a structured handoff packet for human (L2/analyst) review Unlike typical AI copilots, every verdict Lucid produces comes with a fully traceable evidence chain — orchestrated as live, inspectable tasks rather than a single opaque model call — so an analyst or auditor can see exactly which data point led to which decision, not just trust a number. Impact: What normally takes an analyst 10–15 minutes of manual cross-tool investigation happens in seconds, with a defensible, replayable reasoning trail attached, built for the level of accountability that regulated, high-stakes decisions actually require.