FairPrice — AI Marketplace on Vercel + AWS DynamoDB

Nigeria's live escrow marketplace with a Qwen-powered AI agent OS (ZEMA360) — real-time order automation, escrow, payouts, and WhatsApp approvals backed by AWS DynamoDB.

Comment

Inspiration

African informal markets move fast but operate manually. Merchants on FairPrice.ng spend hours on tasks that should be instant — checking inventory, releasing escrow, chasing payouts, notifying buyers. We built ZEMA360 to automate all of it with AI agents backed by a database infrastructure that can scale with the business.

What it does

FairPrice.ng is a live escrow marketplace for Nigeria. ZEMA360 is the autonomous commerce OS built on top of it:

  • Ziva AI — natural language shopping assistant (Alibaba Qwen). Ask "I need an iPhone around ₦1m" and Ziva searches the live catalog, returns product cards with real prices, and starts negotiations
  • Ops Squad — multi-agent pipeline (Inventory → Fulfillment → Finance → Comms) that processes orders end-to-end: checks stock, updates tracking, releases escrow, triggers Paystack payouts, sends WhatsApp updates
  • Human-in-the-loop — large payouts and refunds pause for WhatsApp merchant approval before executing
  • Live dashboard at fairprice.ng/zema360/live — every agent decision logged to AWS DynamoDB, visible in real time with 5-second polling

How we built it

AWS DynamoDB is the core persistence layer for ZEMA360 agent logs. Every agent decision — tool call, approval request, payout trigger — writes to the zema360-agent-logs table (partition key: agent, sort key: ts). The table is auto-created on cold start via ensureTable() and uses PAY_PER_REQUEST billing. The live dashboard reads the last 30 events via a ScanCommand sorted by timestamp descending.

Vercel hosts the entire Next.js 15 App Router frontend. The ZEMA360 events API route (/api/zema360/events) runs as a Vercel Serverless Function with force-dynamic to bypass caching for real-time reads.

Full stack:

  • Frontend: Next.js 15 + Tailwind CSS + Framer Motion on Vercel
  • Database: AWS DynamoDB (agent logs) + Neon Postgres (orders, products, escrow, users)
  • AI: Alibaba Qwen qwen3-max via DashScope — multi-round tool-calling, automatic Gemini fallback
  • Payments: Paystack escrow + payout transfers
  • Comms: WhatsApp Business API
  • Analytics: Novus.ai — 37 track events across the full user journey

Challenges we ran into

  • DynamoDB schema design: Agent logs come in two shapes (ZemaEvent vs AgentLogEntry). Normalising both in the GET handler without a secondary index required careful field mapping
  • Real-time without WebSockets: Achieved near-real-time by polling /api/zema360/events every 5 seconds with AnimatePresence for smooth entry animations — no infrastructure overhead
  • Qwen3 thinking mode: qwen3-max routes output to reasoning_content instead of content for certain queries, causing silent JSON extraction failures. enable_thinking: false fixed it
  • Production safety: DynamoDB replaced only the Firebase log layer — Neon Postgres (core commerce data) was never touched

Accomplishments that we're proud of

  • DynamoDB table live in production (us-east-1), write+read confirmed, zero downtime migration from Firebase
  • The same Next.js API route serves both DynamoDB (primary) and Firebase (fallback) with automatic provider detection
  • Real-time agent dashboard with AnimatePresence running on a live marketplace with real Nigerian merchants and real orders
  • PAY_PER_REQUEST billing — costs literally $0 until traffic arrives, then scales automatically

What we learned

DynamoDB's single-table design shines for append-only event logs. PAY_PER_REQUEST billing with auto-created tables (ensureTable() on cold start) is the fastest path from zero to production. The hardest part wasn't the database — it was normalising two different event shapes from legacy Firebase and new DynamoDB writes without breaking the dashboard.

What's next

  • Move from ScanCommand to a GSI on ts for efficient time-range queries at scale
  • Aurora DSQL for the core commerce tables (orders, escrow) to get distributed SQL at DynamoDB scale
  • Alibaba Function Compute deployment for the Python orchestrator — full Alibaba + AWS + Vercel multi-cloud stack
  • Public enterprise API: POST /api/zema360/process-order with API key auth for other African marketplaces

Built With

  • alibaba-qwen
  • aws-dynamodb
  • framer-motion
  • neon-postgres
  • next.js
  • novus.ai
  • paystack
  • prisma
  • tailwind-css
  • typescript
  • vercel
  • whatsapp-business-api
Share this project:

Updates