RefundIQ

Inspiration

Every small business owner selling on WhatsApp, Instagram, and local e-commerce platforms across Africa faces the same exhausting problem: refund requests in multiple languages, emotional customers, and the constant fear of fraud.

I watched my friend, a Lagos-based merchant, spend 3+ hours daily handling refund complaints in English, Pidgin, French, and Swahili. She had to:

• Manually read every complaint
• Guess if customers were lying
• Translate between languages
• Make refund decisions under pressure
• Deal with angry voice notes

Every existing refund tool was built for Shopify merchants in the US. Nobody was solving this for African merchants.

What it does

RefundIQ is an AI-powered refund management copilot that:

Understands 4 African languages - English, Nigerian Pidgin, French, Swahili
Analyzes voice complaints - Transcribes voice notes and detects emotion/urgency
Detects fraud automatically - Multi-factor risk scoring (0-100)
Provides instant AI decisions - APPROVE/DENY/ESCALATE with reasoning
Responds in customer's language - AI explains decisions in Pidgin, French, etc.
Tracks everything - Real-time analytics, fraud alerts, sentiment trends

Merchants paste a complaint → AI analyzes in 3 seconds → Get recommendation with fraud score

How I built it

Frontend:

• React + TypeScript + Vite
• shadcn/ui components + Tailwind CSS
• Responsive design (desktop + mobile)

Backend:

• Supabase (PostgreSQL database, Row Level Security)
• Supabase Edge Functions (Deno runtime)
• Real-time analytics with SQL aggregations

AI Integration (via MeDo Plugins):

• Speech-to-Text API - Transcribes voice complaints with automatic language detection
• Large Language Model API - Analyzes complaints, detects fraud, generates multilingual responses

Key Technical Features:

• Multi-language AI prompt engineering (4 languages)
• Emotion keyword detection (angry/frustrated/calm)
• Fraud pattern recognition (multilingual keywords)
• Voice recording with MediaRecorder API
• Real-time dashboard with charts (Recharts)
• Role-based authentication (Admin/Manager/Agent)

Challenges I faced

1. Multilingual AI Responses - Getting the LLM to consistently respond in Pidgin/French/Swahili required careful prompt engineering and language mapping

2. Voice Emotion Detection - Building a keyword-based emotion analyzer that works across 4 languages with different expressions of anger/frustration

3. Fraud Detection Logic - Balancing sensitivity (catching real fraud) vs specificity (not flagging legitimate angry customers)

4. Real-time Analytics - Optimizing SQL queries to calculate metrics across thousands of requests without performance issues

5. Mobile Voice Recording - Handling browser permissions and audio format compatibility across devices

Accomplishments

Fully functional multilingual AI - Works seamlessly in 4 languages
Voice feature with emotion detection - Transcribes + analyzes sentiment
Production-ready fraud detection - Multi-factor risk scoring
Complete analytics dashboard - Real-time metrics and charts
25+ test requests - Pre-populated with realistic multilingual data
Mobile responsive - Works perfectly on phones and tablets

What I learned

• AI prompt engineering for multilingual outputs
• Supabase Edge Functions for serverless API integration
• Voice API integration with automatic language detection
• Complex SQL queries for real-time analytics
• Fraud detection algorithms and pattern recognition
• Building for African markets requires cultural context (Pidgin expressions, French formality, Swahili politeness)

What's next

WhatsApp Business API integration - Customers send requests directly via WhatsApp Automated refund processing - Connect to payment gateways for instant refunds
More African languages - Add Yoruba, Igbo, Hausa, Amharic
Customer sentiment tracking - Long-term customer behavior analysis
Bulk import - Upload CSV with multiple refund requests
Email integration - Monitor dedicated email for refund requests

Built With

• api
• deno
• large-language-model-api
• lucide-icons
• mediarecorder
• postgresql
• react
• recharts
• shadcn/ui
• speech-to-text-api
• supabase
• supabase-edge-functions
• tailwind-css
• typescript
• vite