Building an E-commerce AI Agent with AWS Serverless

In this technical deep dive, we’ll design an e-commerce AI agent using an AWS serverless architecture. The agent helps users discover products, get relevant recommendations, ask questions, and perform selected webshop actions.

We’ll walk through the full flow: ingesting product and content data, preparing it for retrieval, combining vector search with full-text search and metadata filters, reranking, and other strategies.

The talk will cover LLM orchestration and tool/function calling, including how an agent can interact with an existing webshop, recommend products, link to relevant blog content, and support cart-related actions through an agent interface.

We’ll also discuss context strategy: what belongs in prompt context, what should be stored as short-term session state or chat history, and what, if anything, should become longer-term memory.

A key part of the talk will focus on safety and reliability challenges: handling anonymous users, applying guardrails, limiting retrieval scope, rate-limiting requests, reducing hallucinations, and keeping responses grounded in product and content data.

Different implementation approaches will be compared. We will analyze the use of Lambda, Bedrock Agents, Bedrock AgentCore, S3 Vectors, and OpenSearch Service. We’ll look at the trade-offs in complexity, flexibility, latency, reliability, and cost.

Finally, we’ll briefly cover evaluation and operations, including model comparison, regression testing, observability, and latency monitoring.

Attendees will leave with practical architecture patterns, implementation trade-offs, and a clear checklist for designing, evaluating, and operating AI agents for e-commerce use cases.