ReturnReady

Inspiration E-commerce reverse logistics is an $850 billion crisis. Most voice shopping assistants fail because listening to an AI read a 20-bullet return policy out loud is terrible UX. We were inspired to build ReturnReady for the Amazon Alexa+ track to move the return intervention upstream—preventing bad purchases before checkout and automating rapid hardware diagnostics post-purchase via a novel split-screen UX (voice for speed, screen for decisions).

What it does ReturnReady is an enterprise-grade, split-screen triage agent.

Pre-Purchase: Matches shopper needs against product risk factors (e.g., warning of a 38% gym return risk due to sweat) to deflect bad buys.

Post-Purchase: Guides users through a 10-second hardware check to resolve issues in the home or route legitimate claims forward.

Security: Bypasses LLM hallucinations using a deterministic Python policy engine that mints tamper-evident, HMAC-SHA256 cryptographically signed claim tokens.

How we built it Protocol & Backend: Built on the Model Context Protocol (MCP spec 2025-11-25) over Streamable HTTP using FastMCP and FastAPI.

Policy Engine: Developed a dependency-light, pure Python deterministic rules engine with 100% Pytest boundary test coverage.

UX & Frontend: Designed a responsive Split-Screen interface using HTML5, Tailwind CSS, and vanilla JS with secure data attribute event delegation.

Standards: Packaged using Kiro Crew open-source standards (.kiro/config.json, SKILL.md).

Challenges we ran into MCP SDK v2 Breaking Changes: The latest package update renamed core classes and shifted transport handling, which we resolved by pinning mcp[cli]<2 and explicitly routing via Streamable HTTP.

Intent Routing & Security: Overcame keyword hijacking in our lightweight router by introducing regex word boundaries, and secured our warranty claims against forgery by replacing static test secrets with strict environment-injected HMAC-SHA256 signing keys (RETURNREADY_SECRET).

Accomplishments that we're proud of Successfully negotiating and displaying live 2025-11-25 MCP protocol telemetry badges directly in the UI.

Engineering an auditable, deterministic policy engine behind an LLM conversational agent, proving that AI can be both conversational and enterprise-compliant.

Achieving a 100% pass rate across our Pytest boundary test suite for complex date and warranty thresholds.

What we learned Voice is exceptional for intent, and screens are essential for structured data. Forcing a user to listen to long blocks of text ruins voice UI; splitting the workload into a 1-sentence spoken verdict via Alexa+ paired with an instant mobile companion UI card unlocks a vastly superior e-commerce experience.

What's next for ReturnReady Integrating live Amazon Bedrock Multimodal capabilities for real-time photo defect verification.

Connecting directly to enterprise carrier shipping APIs (UPS/FedEx) to automate live prepaid label generation.

Expanding the deterministic policy engine to support multi-item bundle returns and automated exchange recommendations.