Graspr: AI Learning Platform with RAG Tutor (Client Project)
Tran Tien Van
The challenge
Students had dense PDFs and recorded lectures but no fast way to turn either into study material they could question.
What we built
A NestJS 10 backend and a Next.js 15 / React 19 frontend in one TypeScript monorepo
RAG on Supabase Postgres with pgvector, combining vector and semantic search through LangChain over Claude and OpenAI
Answers streamed token by token over Socket.io
Live Whisper transcription, with browser-side voice activity detection chunking the audio
Stripe-synced quotas that meter PDF pages, AI prompts and transcription minutes per tier
End-to-end tests with Playwright
Results
RAG answer accuracy above 95%, from a roughly 80% GPT-only baseline
Quiz generation accuracy above 95%, from a 70 to 80% range
Three subscription tiers metered on pages, prompts and transcription minutes
Key decision
Accuracy was a retrieval problem before it was a model problem: the gain came from embedding and search design on pgvector, not from a larger model.
Stack: Next.js 15, React 19, NestJS, TypeScript, Supabase, PostgreSQL, pgvector, LangChain, Claude, OpenAI, Whisper, Socket.io, Stripe, Playwright, AWS.
Full case study with more detail: vandatateam.com/portfolio
Like this project
Posted Nov 5, 2026
Client project for Graspr: RAG AI tutor on Next.js 15 + NestJS. Answer accuracy up from ~80% to 95%+, live Whisper transcription, Stripe usage tiers.
