SunFit

⛬ SunFit is an AI-powered grocery planning assistant that helps users compare costs, choose stores, and adapt their shopping plans. It uses AWS Bedrock AgentCore to host the agent, manage memory, and provide private tracing, while AP4M enables explicitly approved, budget-capped purchases of synthetic price data using test RLUSD—not autonomous grocery purchases.

give a longer description of what we built and what problem we are solving

⛬ What problem are we solving?

A grocery list tells you what you need, but not the best way to get it. Shoppers still have to balance prices, travel time, store availability, budgets, and personal preferences. When something is unavailable or plans change, they often have to work through those decisions again.

SunFit turns that static list into an adaptable shopping journey—while keeping the shopper in control.

What we built

We built a grocery-planning backend that compares shopping options, explains trade-offs, and preserves progress as the user shops. Users can choose a plan, check off items, correct mistakes, and replan what remains without losing decisions they have already confirmed.

The system separates three responsibilities:

 •  AI understands the request. It translates natural-language preferences into a structured proposal.
 •  Deterministic logic makes planning and policy decisions. Budget calculations, constraints, ranking, and spending gates are enforced in code—not
    left to the model.
 •  Dedicated payment services handle approved spending. The model cannot access signing keys or authorize payments.

How AWS Bedrock AgentCore fits in

We deployed the conversational agent on AWS Bedrock AgentCore, integrating Amazon Nova Micro, actor-scoped AgentCore Memory, and private X-Ray tracing. This provides a foundation for contextual conversations and remembering preferences, with tracing designed to exclude prompts, credentials, and shopping content.

How AP4M fits in

We integrated AP4M to demonstrate how an agent can purchase useful information before making a recommendation. In our prototype, the agent can acquire synthetic price lookups using test RLUSD on the XRPL EVM testnet, subject to explicit approval and strict data-spending and gas limits.

Separate Agent and Merchant services run on AWS ECS Fargate, with role-scoped secrets and encrypted payment journals. The payment workflow includes delivery evidence, settlement reconciliation, channel closure, and durable retries designed to prevent duplicate charges.

This is payment for planning inputs—not autonomous grocery checkout.

What we have demonstrated

We tested failures including unauthorized access, excessive spending requests, conflicting retries, lost responses, concurrent workers, insufficient gas allowance, settlement mismatches, and private-content leakage into traces. The project has 178 passing Python tests and 95 passing Node tests, alongside recorded AWS deployment checks.

The current prototype uses synthetic retailer data. A historical local paid sandbox run was verified, but the complete signed-in SunFit → AgentCore → AP4M paid journey remains unverified. The goal is a trustworthy shopping assistant that makes trade-offs clearer and adapts to change—without taking purchasing authority away from the user.