AI-POWERED CAMPUS MARKETPLACE

Project Overview

BuyTree is a full-stack campus marketplace platform built for Nigerian university students — connecting student sellers with campus buyers for peer-to-peer commerce. The platform runs on Express, React, and PostgreSQL. We were brought in to layer a complete AI intelligence system on top of the existing product, integrating six distinct Claude API features that improve the experience for buyers, sellers, and platform admins simultaneously.

All six features share a single backend AI service built on the Anthropic SDK, with a unified architecture: structured inputs, consistent JSON outputs, graceful fallbacks, and rate limiting. AI failures never block core platform actions — every integration is designed to enhance, not gatekeep.

Scope of Work

We designed and implemented six AI features across the full stack — backend services, database schema, API endpoints, and frontend components. Each feature targets a different pain point in the marketplace workflow.

Key deliverables included:

• AI Buyer Support Chatbot: A floating chat widget on all buyer-facing pages powered by Claude Haiku. The bot is order-aware — it fetches the buyer's last 5 orders and injects them as context, so it can answer "where is my order?" without the buyer providing an order number. Scoped strictly to BuyTree: policies, checkout guidance, dispute escalation. Stateless on the server — conversation history lives in React state.

• AI Product Description Writer: Sellers click "Generate with AI" on the product creation form — Claude Sonnet writes a concise, student-market-tuned product description from just the name, price, and category. Sellers accept, edit, or ignore it. Eliminates the blank-page problem that causes thin, unconvincing listings.

• Auto-Categorization: As the seller types a product name, Claude Haiku suggests the correct category from BuyTree's fixed taxonomy in real time (600ms debounce). Returns a confidence level — only high and medium confidence suggestions surface to the seller. Reduces miscategorization and speeds up listing creation.

• Review Authenticity Detection: Every submitted review is asynchronously scored by Claude for authenticity signals — generic text, rating/sentiment mismatch, promotional language, bot patterns. Flagged reviews surface in the admin dashboard with specific flag reasons. Non-blocking: the review saves and returns to the user immediately; AI analysis runs in the background.

• Bypass Detection: A background cron job runs every 6 hours, analyzing order patterns for signs that sellers and buyers are completing transactions off-platform to avoid the 5% fee. Claude evaluates pre-filtered suspicious patterns (cancellation clusters, phone numbers in seller notes, repeat-pair cancellations) and assigns risk levels. High-risk flags trigger an admin email alert.

• Automated Dispute Triage: When a buyer opens a dispute, Claude Sonnet reads the full context — both parties' histories, order status timeline, delivery evidence, and prior disputes — and produces a structured triage report: summary, likely fault, suggested resolution, confidence score, and flags. Admins resolve disputes in a fraction of the time, with the AI recommendation pre-populated for one-click acceptance.