Case Study: Meridian Bank builds a fraud-alert voice agent with Stream

Meridian Bank builds a fraud-alert voice agent in under 16 minutes with Stream

The customer, Meridian Bank, needed a secure, automated system to contact customers about potentially fraudulent transactions without mimicking scammer behavior. They partnered with Stream and used its Vision Agents framework, along with Deepgram for speech-to-text, an LLM, Cartesia for text-to-speech, and Twilio for telephony, to build a compliant voice agent.

Stream's solution was a Python-based voice agent that places outbound calls, identifies the flagged transaction, and asks for a simple yes/no confirmation from the cardholder. The agent's instructions and tooling were strictly designed to never request sensitive data, ensuring legitimacy. The agent runs on Stream's edge network for low latency, achieving a P50 voice round-trip of around 240 milliseconds, and can take real-time actions like clearing a transaction or freezing a card based on the customer's response.

Meridian Bank