Overview

Small businesses often lack the resources to staff a dedicated receptionist around the clock. This project deploys a fully cloud-hosted AI conversational agent that handles inbound customer interactions in real time using speech-to-speech AI — no local servers, no on-premises infrastructure required.

The solution was adapted from the AWS Nova Sonic CDK sample and deployed for a small retail hardware store. The agent acts as a digital receptionist: answering product and service questions, guiding customers through common requests, and responding naturally in real time.

The Challenge

Traditional IVR systems feel robotic and frustrate customers. The goal was to deploy something that could hold a natural, context-aware conversation tailored to the store’s specific inventory and service flow. The solution also needed to be fully cloud-native and cost-effective at small business scale.

The key customization work happened at the data layer — configuring the DynamoDB schema with the right indexes and interaction variables to capture meaningful context from each customer conversation, enabling the agent to respond accurately to store-specific queries.

Architecture

The architecture is entirely serverless and managed:

  • Frontend — Static web interface hosted on Amazon S3, distributed via Amazon CloudFront
  • Backend — Python application running on Amazon ECS, handling WebSocket connections and tool orchestration
  • AI Model — Amazon Nova Sonic via Amazon Bedrock, processing speech-to-speech interactions in real time
  • Data Layer — Amazon DynamoDB with custom indexes and interaction variables tailored to the client’s product catalog and customer service flow
  • Authentication — Amazon Cognito securing session access
  • Connectivity — Network Load Balancer managing WebSocket traffic between frontend and backend

When a customer initiates a session, audio is streamed bidirectionally over WebSocket. Nova Sonic processes the incoming speech, generates a response or triggers a tool call against the knowledge base, and streams the audio response back in real time. Infrastructure is defined entirely in AWS CDK.

Outcome

The client saw measurable improvements in customer satisfaction and a more streamlined client attention flow. Routine inquiries that previously required staff intervention are now handled automatically, freeing the team to focus on higher-value in-store interactions. The agent operates beyond business hours, extending the store’s customer service availability without additional staffing cost.