Seller Shield

An AI agent team that protects small Amazon sellers - it catches account-health risk early and drafts the evidence-backed appeal or SAFE-T claim for human review.

Comment

Inspiration

Amazon's seller enforcement is now almost fully automated: a compliance flag can suppress a listing or suspend an account within minutes, with human review only happening after the seller appeals. At the same time, return fraud has climbed to roughly 14% of all retail returns (up from ~5% in 2018), and Amazon's SAFE-T reimbursement program caps recovery around 50% for damaged returns and doesn't guarantee payout. Small third-party sellers — often solo operators — have no compliance team to catch this in time or to fight back with properly assembled evidence. We built Seller Shield to be that team.

What it does

Seller Shield is a team of four AI agents, coordinated by an LLM-routing orchestrator, that does the actual paperwork instead of just explaining policy:

  • Monitor watches account-health metrics and return data, detecting suspension risk, return-fraud spikes from a single buyer, and ODR/OTDR drift — before an enforcement action hits.
  • Evidence gathers supporting documents (invoices, shipping proof, product images) and retrieves the applicable Amazon policy clause from a sourced corpus — never a guessed citation.
  • Drafting writes the SAFE-T reimbursement claim or suspension appeal, citing the evidence and policy clause. Every draft is for human review and is never auto-submitted to Amazon.
  • Escalation tracks the claim's lifecycle (submitted → pending → rejected → re-drafted → approved → resolved) and recommends the next action at each step.

A React dashboard surfaces account health, active risk events, editable claim drafts, and a claim-status timeline.

How we built it

  • Strands Agents SDK on Amazon Bedrock (Claude Sonnet 4.6), using the Agents-as-Tools pattern.
  • Every agent has a deterministic, unit-tested core (threshold detection, evidence retrieval, claim templating, the escalation state machine) with an LLM layer on top. This means the whole pipeline runs and is fully tested without live model access — 138 passing tests — while the top-level orchestrator still does genuine LLM routing over the tools on Bedrock.
  • FastAPI backend exposing structured, deterministic endpoints; React + Vite dashboard.

Challenges we ran into

  • Keeping every number trustworthy — detection is deterministic (named, tunable thresholds in code), so the model never invents a metric.
  • Passing typed state (RiskEvent, EvidenceBundle, ClaimStatus) across LLM tool calls, solved with an in-process session store keyed by event/claim id.
  • New-account AWS onboarding: enabling Bedrock model access, the Anthropic use-case form, and the Marketplace subscription.

Accomplishments that we're proud of

  • A working end-to-end pipeline and dashboard, 138 passing tests, and verified live LLM routing on Amazon Bedrock.
  • An honest design: Seller Shield drafts the paperwork, but a human always reviews and it never submits to Amazon on the seller's behalf.

What we learned

Splitting a deterministic, testable core from the LLM narration/routing layer makes an agent system reliable to demo, cheaper to run, and trustworthy — while still showcasing real agentic orchestration.

What's next for Seller Shield

Real Amazon SP-API integration for live seller data, a Bedrock Knowledge Base for a fully sourced, up-to-date policy corpus, and optional Amazon Bedrock AgentCore deployment.

Links

Built With

  • amazon-bedrock
  • claude
  • fastapi
  • python
  • react
  • strands-agents
  • vite
Share this project:

Updates