Deploy a multimodal WhatsApp ordering assistant with Amazon Bedrock AgentCore
aws ·
The article describes deploying a multimodal WhatsApp ordering assistant for quick-service restaurants, integrating text, voice notes, and calls through one AI agent. The solution uses WhatsApp Business as the front door, Amazon Bedrock AgentCore for hosting three channel-specific agent runtimes, and Amazon Nova 2 Lite for text and Nova 2 Sonic for speech-to-speech voice interactions. A shared AgentCore memory keyed by hashed customer ID provides cross-channel continuity, while a managed MCP server (AgentCore Gateway) exposes backend ordering tools to the agents. The AWS architecture includes API Gateway, Lambda, SQS, DynamoDB, Location Service, Kinesis Video Streams for TURN relay, and a VPC for voice calls, all deployed via AWS CDK. Each channel (text, voice note, voice call) follows the same flow: webhook ingestion, queue processing, agent runtime invocation, tool calls, and reply delivery. The post also outlines prerequisites including AWS account setup, model access, and Meta WhatsApp configuration. Overall, the design decouples channels from backend logic, allowing restaurants to manage orders across messaging platforms seamlessly.
aws