Serverless Real-time Voice AI on AWS: A Pattern for Enterprise Sales Coaching

This article presents a serverless architecture pattern for building real-time voice AI applications, demonstrated through an enterprise sales coaching solution using AWS AppSync Events, Amazon Nova Sonic, and Amazon Bedrock AgentCore.

• AWS AppSync Events provides serverless WebSocket API for bidirectional audio streaming between web clients and backend containers
• Two-channel pub/sub pattern: upstream channel for client-to-container audio, downstream channel for container-to-client responses
• Cost model of $1.00 per million event operations with zero idle costs, approximately $0.007 per five-minute coaching session
• Production deployment serves field sales representatives with personalized product recommendations and objection handling
• Critical configuration: use API-specific DNS endpoint for WebSocket subscriptions; generic regional URL fails silently
• Deployment checklist includes implementing GET /ping health checks, disabling HTTP/2 timeouts, and session-level telemetry logging
• Compound failure pattern discovered: missing health checks combined with wrong AppSync endpoint caused all sessions to terminate at ~112 seconds

The serverless audio transport pattern eliminates WebSocket server management and provides automatic scaling, making real-time voice AI economically viable for enterprise applications beyond sales coaching.