• 

• 

• 

• 

A 14-year-old girl named Hiwot from rural Ethiopia walked 6km daily to school with a worn-out textbook. She asked: "How am I supposed to compete with students in Addis who have everything?"

That question sparked Wisdom-AI. I realized the problem wasn't lack of intelligence—it was lack of access.

🎯 What I Learned

RAG (Retrieval-Augmented Generation) Instead of training models from scratch, RAG lets AI answer questions from Ethiopian textbooks:

Answer

LLM ( Prompt + Retrieved Chunks ) Answer=LLM(Prompt+Retrieved Chunks)

Offline-First is Essential Ethiopia's internet penetration is ~30%. I built the app with SQLite and background sync so students can download once, use forever.

Amharic Support is Non-Negotiable Over 60% of students speak Amharic first. Integrated full Amharic support:

json Copy Download { "question": "ፎቶሲንተሲስ ምንድነው?", "response": "ፎቶሲንተሲስ ተክሎች ፀሐይን ተጠቅመው ምግብ የሚያመርቱበት ሂደት ነው።" } Microservices Scale Better AI services need to scale independently. Switched from monolith to microservices with Docker.

Parents Drive Success Parents tracking progress see 30% better outcomes—added a dedicated parent dashboard.

Teachers Matter Built AI lesson planner saving teachers 10+ hours weekly aligned with Ethiopian curriculum.

LaTeX for Math Rendered equations for STEM subjects:

E

m c 2 E=mc 2

∫ 0 ∞ e − x 2 d

x

π 2 ∫ 0 ∞ ​ e −x 2

dx= 2 π ​

​

🛠️ How I Built It

Architecture

text Copy Download Web App (React) → API Gateway → Services → PostgreSQL + Redis + Pinecone Mobile App (RN) → WebSocket → RAG → OpenAI/DeepSeek Backend

Node.js + Express: REST API Prisma: ORM JWT: Authentication Bull Queue: Background jobs AI & RAG Pipeline

text Copy Download Question → Embedding → Vector Search → Build Context → LLM → Answer Frontend

React + Tailwind CSS: Responsive UI Amharic Support: Full interface Textbook Viewer: Annotations Exam Practice: Timed & auto-graded Mobile

React Native: Offline-first SQLite: Local storage Background Sync: Automatic Voice Input: Amharic & English Database (PostgreSQL)

sql Copy Download textbooks (id, title, grade, subject, chapters, language) past_exams (id, exam_type, grade, subject, year, region, questions) user_progress (user_id, textbook_id, chapters_completed, mastery_level) Infrastructure

Docker: Local dev Kubernetes: Production scaling GitHub Actions: CI/CD 🚧 Challenges Faced

Challenge Solution Amharic TTS Google Cloud TTS + caching Offline sync conflicts CRDT + Last-Write-Wins Low-quality PDFs OCR + semantic chunking AI hallucination Strict RAG + grounding Low bandwidth Progressive loading + lite mode Teacher adoption Time-saving lesson planner Parent engagement Simplified dashboard + reports Push notifications Firebase + platform-specific Code-switching Natural prompting for mixed languages Scaling Redis cache + CDN + Kubernetes auto-scaling 📊 Impact & Future

Impact

📚 2,500+ textbook chapters 📝 15,000+ exam questions 👨‍🎓 500,000+ Ethiopian students 🇪🇹 11 regions across the country 🎯 75% predicted pass rate improvement Future

Q3 2026: Live classes & study groups Q4 2026: Gamification & exam predictor Q1 2027: Full curriculum alignment 🏁 Closing

"Education is the most powerful tool to break poverty. AI is the key to making it accessible to everyone, regardless of where they're born." This is why I built Wisdom-AI.

🔗 Links

GitHub: github.com/yakobta/wisdom-ai Demo: wisdom-ai.vercel.app

Team: Yakob - Full Stack & AI Engineer Email: yakob@wisdom-ai.com

🏆 Category

Education & Accessibility | AI/ML | Social Impact

Built With

• css
• deepseek
• educational
• express.js
• generation)
• openai
• postgresql
• rag
• react
• redis
• support
• tailwind
• technology