Vanguard Pay Fraud Resolution

The problem

When Vanguard Pay's Enterprise Fraud Management Solution flags suspicious behaviour, the account is restricted in milliseconds. But the resolution process that follows is manual and asynchronous. Customers are left without immediate context, support queues swell, and anti-fraud analysts spend their time gathering basic facts instead of deciding hard cases.

Before the agent

A fraud flag triggered an internal ticket queue. The customer found out when a transaction failed, then called support. Analysts manually gathered transaction intent, verified identity, and decided whether to unblock. The delay between detection and resolution could be hours.

What I built

An AI + automation support agent acts as a Tier 1 triage layer. The moment a fraud flag is raised, the customer receives a secure, time-sensitive link. The agent asks one context-aware security question at a time, watches for coercion or inconsistency, and produces a structured JSON summary for the human analyst who executes the final unblock.

• Detection and alerting. EFMS triggers an n8n webhook that creates an incident record, notifies the anti-fraud team, and sends a secure verification link to the customer.
• Conversational triage. A Next.js chat UI hosts an AI agent that asks dynamic security questions based on the transaction context, with strict guardrails.
• Structured handoff. On completion, n8n generates a JSON summary and updates the incident record in Supabase for analyst review.
• Analyst dashboard. A Next.js admin view lets anti-fraud analysts review the triage summary and execute the final unblock command.
• Clear tier separation. Tier 1 autonomous triage gathers facts; Tier 2 human authorization makes the unblock decision.

The outcome

Built as a capstone project, the system closes the gap between real-time detection and human resolution. The intended outcome is less customer distress, fewer support escalations, and analysts spending their time on decisions rather than data collection.

Stack

Next improvements

• Biometric verification. Add device binding or push-to-app confirmation as an additional verification factor.
• Case analytics. Track resolution time, escalation rate, and false-positive trends.
• Integration with EFMS. Move from simulated webhook to direct API integration with the production fraud management system.