ReRoute

Inspiration

The average American household has over $3,000 in unused items, yet most people never recover that value because the process is overwhelming. Return it? Trade it in? Sell on eBay? Repair it first? Each option requires research, pricing, listing, and negotiating — multiplied by every item.

We wanted to make it as easy as recording a quick video. That's ReRoute: Record once. We handle the rest. The Fetch.ai uAgents framework was the perfect fit — specialized AI agents that bid against each other to find the best "route" for every item, like buyers competing at an auction.

What it does

ReRoute turns a single narrated video of your unused items into optimized listings across multiple marketplaces.

Record — Point your phone at your items and narrate their condition. One video, all items.
Identify — AI fuses your speech with visual analysis to identify each item, infer specs, and detect defects.
Bid — Five agents simultaneously evaluate every item: Return, Trade-In, Resale, Repair-then-Sell, and Bundle.
Decide — A Route Decider scores all bids on value, confidence, effort, and speed, then picks the winner.
Optimize — Listing images are auto-cropped, enhanced, and background-removed.
Execute — Listings publish concurrently to eBay, Mercari, and more with one click.
Manage — A unified inbox aggregates buyer conversations across platforms with AI reply suggestions and buyer seriousness scoring.

How we built it

ReRoute runs on a 9-agent architecture powered by the Fetch.ai uAgents framework, all orchestrated through a Bureau.

AI Engine: Gemini 3.1 Pro for multi-modal fusion (video + speech), comp search, listing generation, and reply suggestions. ASI:One powers the Concierge chatbot on Agentverse.
Agents: Typed Pydantic protocol messages between agents — ConditionFusionAgent fans out ItemCards to 5 bidding agents, collects bids, and forwards to RouteDeciderAgent. Low-confidence bids trigger delegation for re-evaluation.
Backend: FastAPI + WebSockets for real-time updates. Event-driven JobStore broadcasts state changes to connected clients.
Image Pipeline: OpenCV for sharpness/brightness scoring, Pillow for auto-crop and exposure normalization, rembg for background removal.
Live APIs: eBay Browse + Sell APIs, Apple Trade-In API, and Amazon Product Advertising API for real market data.
Frontend: React + Vite dashboard with live agent visualization, plus a mobile-optimized phone capture UI.

Challenges we ran into

Multi-modal fusion — Getting Gemini to reliably distinguish items in a continuous video required careful prompt engineering to cross-reference frames with transcript timestamps.
Agent coordination — Five agents bidding simultaneously with wildly different response times. We built timeout-based collection with partial result handling.
Rate limits — Concurrent Gemini calls across 9 agents hit limits fast. We built round-robin key rotation supporting up to 10 API keys.
eBay API — Their multi-step Sell API flow (inventory → offer → publish) with strict schema validation required extensive sandbox experience.

Accomplishments that we're proud of

A fully functional 9-agent system with genuine competitive bidding — not simple task delegation, but independent reasoning and collaboration.
Real multi-modal AI fusion that catches defects mentioned in speech but not visible on camera, and vice versa.
Live production API integrations (eBay, Apple Trade-In, Amazon) returning real market data and publishing real listings.
End-to-end pipeline from video upload to published listing in a single seamless flow with real-time WebSocket visualization.

What we learned

Competitive bidding > fixed decision trees — Having agents bid makes the system extensible. Adding a new route is just adding a new agent.
Confidence scoring is critical — A $200 estimate at 40% confidence is worse than $150 at 95%. Our composite score (value × confidence × effort × speed) produced much better recommendations.
Fetch.ai uAgents scales well — Bureau, typed protocols, and Agentverse gave us production-grade agent communication without building messaging infra from scratch.
Multi-modal AI is more than the sum of its parts — Combining vision and speech enabled cross-confirmation (seeing shrink wrap + hearing "never opened") that neither modality could achieve alone.

What's next for ReRoute

More marketplaces — Facebook Marketplace, Depop, Poshmark, and OfferUp adapters.
Receipt ingestion — Pull purchase history from email to pre-fill return windows, warranties, and original prices.
Route learning — Feed actual sale outcomes back into scoring so ReRoute improves with every item.
Native mobile app — Replace the web capture UI with iOS/Android for smoother recording and push notifications.