ReVival — AI-Powered Circular Commerce

Amazon HackOn Season 6 | Second-Life Commerce | Live demo: https://revival-rose.vercel.app

ReVival is an end-to-end circular-commerce system that routes returned and seller-listed products back into the market instead of landfill. It uses a 7-agent pipeline to grade product condition from photos/video, route items to the correct disposition, recommend pricing, match items to low-return-risk buyers, generate trust passports, and prevent repeat listing errors.

Why this project matters

Returns are expensive, wasteful, and often poorly routed. ReVival treats every returned item as a recoverable asset by combining AI vision, deterministic business rules, explainable buyer matching, and sustainability incentives.

My contribution

I owned the supply-side AI + AWS backend:

• Built the product return and community-listing pipeline in FastAPI.
• Implemented the AI vision grading agent using AWS Bedrock vision models.
• Added photo/video-based condition grading with evidence, defect detection, size/color mismatch detection, and confidence buckets.
• Built the AWS data layer using DynamoDB, S3, and boto3.
• Added Titan multimodal embedding caching to reduce redundant Bedrock calls on visually similar products.
• Connected grading output to disposition, pricing, trust-passport generation, and buyer recommendation flows.

System at a glance

Area	Implementation
Backend	FastAPI, Python 3.12, boto3
Frontend	Next.js, TypeScript
AI stack	AWS Bedrock, Qwen3-VL / Claude-compatible vision flow, Mistral/Claude-compatible text flow
Storage	DynamoDB + S3
Agents	7 total: grading, disposition, pricing, matching, green credits, passport, prevention
Caching	Titan multimodal image embeddings + deterministic text cache
Demo data	30 buyers, 15 seed items, 10 Indian cities, 13 categories

Pipeline

Seller uploads return/listing media
        ↓
AI Vision Grading Agent
        ↓
Disposition Agent
        ↓
Pricing Agent
        ↓
BuyerInterestIndex candidate retrieval
        ↓
LLM-assisted Matching Agent
        ↓
Green Credits calculation
        ↓
Trust Passport Agent
        ↓
Prevention Agent
        ↓
Item listed in buyer recommendation feed

Core features

1. AI product grading

The grading agent accepts photos or videos and returns:

• condition grade: A, B, C, D, or REVIEW
• confidence bucket
• detected defects
• wear level
• functional-status signal
• size/color mismatch detection
• evidence fields for auditability
• image-cache metadata

Guardrails are applied after the LLM response:

• low confidence → REVIEW
• unsafe or hygiene-blocked products → D or REVIEW
• electronics with unknown functionality → REVIEW
• not-working products → D

2. Similarity cache for image grading

To reduce repeated Bedrock calls, product images are canonicalized, embedded through Titan multimodal embeddings, and looked up in an image-vector cache before calling the grading model again.

3. Buyer matching

The matching flow uses a two-stage ranking system:

1. Deterministic retrieval from BuyerInterestIndex.
2. LLM-assisted reranking using return-reason neutralization and recurrence risk.

The final risk score exposes factor-level explanations, including:

• buyer return rate
• size incompatibility
• condition intolerance
• brand affinity
• return-reason recurrence
• return-reason neutralization
• eco-credit boost

4. Trust Passport

Each resale item gets a human-readable condition report with:

• condition summary
• why the item was returned
• buyer reassurance
• AI verification explanation
• S3-hosted passport page

5. Prevention loop

If the system detects wrong size/color attributes, it writes listing flags so the original product page can warn future buyers and reduce repeat returns.

API overview

Method	Endpoint	Purpose
POST	/returns	Process a returned product through the full pipeline
POST	/community-list	Process a peer-to-peer seller listing
POST	/grade-preview	Grade media before listing submission
GET	/buyers/{buyer_id}/recommendations	Personalized resale recommendations
GET	/items/{item_id}	Item detail with grading evidence and buyer-risk breakdown
GET	/items/{item_id}/passport	Trust Passport data + S3 URL
GET	/ops/items	Operations dashboard feed
POST	/credits/redeem	Redeem green credits at checkout

Local setup

Backend

cd backend
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
cp ../.env.example ../.env
python seed/seed.py
uvicorn main:app --reload --port 8000

Frontend

cd frontend
npm install
npm run dev

Create frontend/.env.local:

NEXT_PUBLIC_API_BASE_URL=http://localhost:8000
NEXT_PUBLIC_DEMO_BUYER_ID=BUY-001

Environment variables

Use .env.example as the source of truth. Do not commit real AWS bucket names, access keys, or account-specific infrastructure values.

Required categories:

• AWS region/profile
• Bedrock model IDs
• DynamoDB table prefix
• S3 bucket names
• CORS origins
• demo mode flag

Repository structure

ReVival/
├── backend/
│   ├── main.py
│   ├── orchestrator.py
│   ├── cache.py
│   ├── agents/
│   │   ├── grading.py
│   │   ├── disposition.py
│   │   ├── pricing.py
│   │   ├── matching.py
│   │   ├── green_credits.py
│   │   ├── passport.py
│   │   └── prevention.py
│   ├── db/
│   └── seed/
└── frontend/
    ├── pages/
    ├── components/
    ├── lib/
    └── data/

Numbers at a glance

Metric	Value
Backend agents	7
LLM/vision-backed agents	3
Deterministic agents	4
DynamoDB tables	7
S3-backed artifact types	photos + trust passports
Frontend pages	12
React components	9
Seed buyers	30
Seed items	15
Risk-score factors	7

Demo checklist

Before sharing this repo, add:

• product upload/grading GIF
• buyer recommendation screenshot
• Trust Passport screenshot
• ops dashboard screenshot
• architecture diagram
• short demo video link

Status

Hackathon prototype with working local/demo flows. The next step is deployment hardening: secrets management, CI checks, structured integration tests, and public demo instrumentation.