Skip to main content
Back to case studies
Graspr AI logoClient projectLive

Education · AI Agents

Graspr AI Learning Platform

Built by Van Data Team for Graspr

Full-stack learning platform that turns documents, audio, and video into document Q&A, generated quizzes, flashcards, summaries, and chapter detection behind subscription tiers.

RAG answer accuracy, from about 80%
95%+
students on the platform, per graspr.ai
130K+
graspr.ai
Graspr home page
Graspr on a phone

Measured outcomes

  1. Outcome 01

    RAG answer accuracy above 95%, from a roughly 80% GPT-only baseline

  2. Outcome 02

    Quiz generation accuracy above 95%, from a 70 to 80% range

  3. Outcome 03

    Three subscription tiers metered on pages, prompts, and transcription minutes

Challenge pressure

Students had dense PDFs and recorded lectures but no fast way to turn either into study material they could question.

The situation

Course material arrives as long PDFs and hour-long recordings. Reading them is not the bottleneck; interrogating them is. A student cannot ask a lecture a question, and cutting one into quiz material by hand costs more time than the revision it was meant to support.

What we built

A NestJS 10 backend and a Next.js 15 with React 19 frontend in one monorepo. Answers stream token by token over Socket.io. Browser-side voice activity detection chunks audio for live Whisper transcription. Retrieval runs on Supabase Postgres with pgvector, combining vector and semantic search through LangChain over Anthropic Claude and OpenAI. Stripe-synced quotas meter PDF pages, AI prompts, and transcription minutes per tier, and Playwright covers the flows end to end.

What changed

Retrieval accuracy on generated answers moved from roughly 80% on the earlier GPT-only baseline to above 95%. Quiz generation rose into the same band from a previous 70 to 80% range.

Constraints and decisions

Accuracy here was a retrieval problem before it was a model problem: the gain came from embedding and search design on pgvector rather than from a larger model. Metering had to sit alongside usage rather than after it, because quotas priced in pages, prompts, and minutes are only defensible if they cannot drift from what Stripe billed.

Technology stack

NestJS 10Next.js 15React 19SupabasepgvectorLangChainSocket.ioStripeAWS

Share this

Need a similar system?

If this case study maps to your current workflow, the next useful step is a scoped project conversation with constraints, timeline, and integration needs.

Related case studies