Inspiration

Every Pakistani university student has a folder full of lecture slides they barely understand. They can't afford tutors. Generic AI tools like ChatGPT hallucinate — they answer from the internet, not the student's actual course material. I wanted to build something that actually helps.

What it does

SabaqAI lets students upload their lecture PDFs and PowerPoint slides, then chat with an AI tutor that answers questions sourced directly from their own materials — in Roman Urdu or English. Every answer includes a citation showing which file it came from. No hallucination. No guessing.

How I built it

Built entirely on AWS using Amazon Nova Lite as the core LLM, accessed via the Bedrock RetrieveAndGenerateCommand with an APAC cross-region inference profile. Lecture files are stored in S3, indexed by a Bedrock Knowledge Base using Titan Embeddings V2, and stored as vectors in Amazon OpenSearch Serverless. The frontend is Next.js 16 with Tailwind CSS deployed on Vercel. Auth is NextAuth.js, database is MongoDB Atlas.

Key technical challenges solved:

  • File size limits: Used S3 presigned URLs so files upload directly from browser to S3, bypassing Vercel's 4.5MB limit entirely
  • PPT support: Integrated CloudConvert API to convert PPTX to PDF server-side before indexing
  • APAC inference: Used the APAC cross-region inference profile for Nova Lite since on-demand throughput isn't available directly in ap-south-1

What I learned

How to build a production RAG pipeline on AWS from scratch. The power of Amazon Nova's multilingual understanding — it handles Roman Urdu natively without any special configuration. How to architect a Next.js app with server components doing data fetching and client components handling interactivity.

What's next

Voice support with Amazon Nova Sonic, automated quiz generation with Nova Act, and expansion to universities across Pakistan and South Asia.

Built With

  • amazon-nova-lite
  • amazon-opensearch-serverless
  • amazon-web-services
  • aws-bedrock
  • cloudconvert
  • mongodb-atlas
  • next.js
  • next.js-16
  • nextauth.js
  • tailwind-css
  • titan-embeddings-v2
  • vercel
Share this project:

Updates