← Back to portfolio

VTEX Logistic

Live

AI agent that answers in natural language about an API, down to the smallest detail of each endpoint.

Product

I built VTEX Logistic because in my experience API catalogs organize documentation well, but fall short at search — the filters are technical and deterministic, and if the term doesn't match exactly how it was indexed, the answer doesn't show up even if it exists.

  1. Natural language searchYou ask the way you'd talk to a colleague: what you need to create a shipping policy, what parameters an endpoint requires, how to check a SKU's inventory.
  2. Minimum detail or aggregate viewYou drill down to a specific field of an endpoint or ask something broader, like how many operations an entire category has.
  3. Ready-to-use responsesEach response comes with parameter tables, code blocks (curl examples), and copyable request/response JSON, instead of a paragraph you have to interpret.
  4. Real-time streamingYou watch the response take shape live, with the text and UI blocks appearing in the same order the agent generates them.
VTEX Logistic: home screen with shortcut questions
Home screen with domain shortcut questions.
VTEX Logistic: response with a table of required parameters
Response with a table of required parameters.
VTEX Logistic: response with a sample JSON block
Sample JSON and required permissions.

Technical

  • I built the backend with AWS Bedrock AgentCore Runtime, with Native Function Calling.
  • The model picks the UI component (table, JSON, code, list) for each part of its response.
  • I built the frontend in static Next.js, with conversation memory across turns via AgentCore Memory.
  • Cognito validates the user token before the agent responds.