Neuro-Study Ai

"NeuroStudy AI — Transform any PDF into a complete learning experience with 5 specialized AI agents. One PDF. Infinite Learning."

Comment

Inspiration

As a student, I constantly struggled with the overwhelming amount of study material — PDFs full of notes that took hours to process manually. I wanted a smarter way to study: one that could read a document and instantly generate notes, quizzes, doubt answers, study plans, and revision strategies. That frustration became the foundation of NeuroStudy AI — a system where multiple specialized AI agents work together to transform any PDF into a complete learning experience.

What it does

NeuroStudy AI is a multi-agent AI study assistant that takes a PDF as input and unleashes 5 specialized agents:

  • Notes Agent — Extracts key points, definitions, summaries, and topic tags from any PDF
  • Quiz Agent — Auto-generates MCQs with multiple difficulty levels and detailed explanations
  • Doubt Agent — Chat-style Q&A where you can ask anything about your document and get instant AI answers
  • Planner Agent — Generates a personalized day-by-day study schedule based on the document content
  • Revision Agent — Identifies weak/difficult topics and creates a focused revision plan with practice questions

All agents share the same uploaded document and are accessible from a unified dashboard.

How we built it

Frontend: Next.js 16 (App Router) + Tailwind CSS — dark futuristic cyber-tech UI with per-agent accent colors, glass morphism cards, and animated backgrounds. Deployed on Vercel.

Backend: FastAPI (Python) with 5 independent agent modules, each with its own router, model, and prompt engineering logic. Deployed on Render.

AI: Groq API using llama-3.3-70b-versatile — chosen for its speed and reliability. Each agent sends a carefully engineered prompt and parses structured JSON from the model response.

Database: MongoDB Atlas (py mongo, synchronous) — stores uploaded documents and chat history for the Doubt Agent.

Auth: Firebase Authentication with Google Sign-In.

Architecture: Each agent follows the same pattern — agents/X_agent.py → routers/X.py → registered in main.py → frontend page at app/dashboard/X/page.tsx. PDF text is extracted on upload and stored in MongoDB, then passed to whichever agent the user activates.

Challenges we ran into

  • Groq API key compatibility — Indian Google accounts generate OAuth-based AQ. format keys incompatible with Gemini's direct API key auth. Switched to Groq as the AI provider, wrapping it in a call_gemini() function to maintain consistent interfaces across all agents.
  • JSON parsing failures — Every agent prompts the model to return pure JSON. Groq occasionally returned empty responses or added markdown fences around the JSON, causing JSONDecodeError. Fixed by adding empty-response guards, stripping markdown fences, and surfacing raw model output on failure for debugging.
  • Firebase auth loop on deployment — sign In With Redirect failed silently on Vercel due to cross-origin cookie restrictions in modern browsers. Fixed by switching to sign In With Popup and removing a broken middleware that checked for a cookie Firebase never sets.
  • Vercel build failure — JSX.Element type reference failed in Vercel's strict TypeScript build environment. Fixed by replacing it with the properly imported ReactElement from React's type definitions.
  • MongoDB sync vs async — Early attempts to use await on py mongo calls caused silent failures. Enforced synchronous-only database access across all agents.

Accomplishments that we're proud of

  • Built and deployed a fully functional multi-agent AI system end-to-end in a hackathon timeframe
  • Designed a consistent, premium cyber-tech UI system across 5 agent pages with unique per-agent color themes — all without Framer Motion or any heavy animation libraries
  • Each agent produces genuinely useful, structured output (not just raw text) — MCQs with options and explanations, day-by-day plans with topics and time estimates, revision plans with practice questions
  • Successfully debugged and resolved 5+ distinct deployment/auth/API issues across frontend, backend, and third-party services
  • The Doubt Agent maintains conversation history per document, enabling true multi-turn chat with context

What we learned

  • Multi-agent architecture requires careful prompt engineering per agent — a generic prompt doesn't work; each agent needs specific output schemas and rules to produce reliable structured JSON
  • sign In With Redirect is unreliable in deployed environments; sign In With Popup is the correct choice for production Firebase auth
  • Vercel's production TypeScript build is significantly stricter than local development — types that work locally can fail in CI
  • Groq's LLaMA 3.3 70B is fast and capable for structured generation tasks, but requires explicit max_tokens and robust error handling to avoid silent truncation
  • Building a real full-stack AI product reveals layers of complexity that tutorials don't cover — environment variables, CORS, cookie behavior, cross-origin auth, and API rate limits all matter in production

What's next for Neuro-Study Ai

  • Voice input for Doubt Agent — ask questions by speaking instead of typing
  • Multi-document support — let agents cross-reference multiple PDFs simultaneously
  • Progress tracking — track quiz scores over time and adapt the revision plan based on performance
  • Collaborative study rooms — share a document with friends and study together with shared notes and quizzes
  • Mobile app — React Native version for studying on the go
  • Support for more file types — Word documents, PowerPoint slides, and YouTube video transcripts I’ve made many changes to the AI’s UI.

Built With

Share this project:

Updates