NovaBuddy

• 

  Quiz

• 

  Landing

• 

  Summary

• 

About the Project

Inspiration

As a computer science student juggling dense lecture slides, long readings, and endless notes, I often felt overwhelmed trying to review everything before exams. I wanted an AI that could instantly digest uploaded materials, pull out the most important ideas, create useful practice questions, and let me ask follow-up questions without losing the full context. When I learned about the Amazon Nova AI Hackathon and saw Nova 2 Lite — a fast, affordable multimodal reasoning model with a huge 1M token context window and strong document/image understanding — it felt like the perfect tool to solve this real student problem. I set out to build something that could make learning more efficient and accessible, especially for students who can't afford premium tutoring services.

What it does

NovaBuddy is my personal AI study companion that transforms uploaded lecture materials into clear, actionable study tools in seconds.

• Upload PDFs, slide decks, images, diagrams, or screenshots directly from the browser
  
• Nova 2 Lite analyzes the content (text + visuals like charts, equations, diagrams)
  
• Automatically generates clean markdown summaries, key concepts, and 8–12 high-quality quiz questions (multiple-choice + short answer)
  
• Chat naturally with your lecture — ask things like “Explain this formula from slide 4” or “What’s the difference between concept A and B?” — and get accurate, context-aware answers that remember the entire uploaded material thanks to the large context window
  
• Everything stays private and secure per user (AWS Cognito authentication)

It saves hours of manual note-taking and helps students deeply understand (not just skim) their course content.

How I built it

I created NovaBuddy as a fully serverless web application on AWS, optimized for speed, low cost, and hackathon constraints:

• Backend (deployed entirely with Terraform):

  • Amazon Cognito for secure email/password sign-up/login and JWT authorization
    
  • HTTP API Gateway with Cognito authorizer
    
  • Python 3.12 Lambda functions calling the Bedrock Converse API for amazon.nova-2-lite-v1:0 (used for summarization, quiz generation, and conversational reasoning)
    
  • S3 buckets with presigned PUT URLs for direct, secure uploads from the browser
    
  • Single-table DynamoDB design to store per-user lecture metadata and AI-generated outputs
    

• Frontend — Next.js 16 (App Router) with shadcn/ui components + Tailwind CSS (dark cosmic theme)

  • Upload modal → requests presigned URL → direct S3 upload → triggers processing
    
  • Dashboard showing lecture cards with quick previews
    
  • Clean, responsive lecture detail page with sections for summary, key concepts, quiz, and interactive chat
    
  • Markdown-rendered responses and message history in the chat interface

The whole system connects via secure REST APIs. Total build time: roughly 48 focused hours, mostly spent on prompt engineering, IAM debugging, and UI polish.

Challenges I ran into

• Prompt engineering took many tries — Getting consistent, well-structured JSON quizzes and nicely formatted markdown summaries required careful temperature settings, few-shot examples, and explicit output format instructions. Early attempts sometimes mixed up diagram explanations or produced messy formatting.
  
• Handling complex document visuals — While Nova 2 Lite understands images and layouts natively, extracting and reasoning about intricate charts/equations in PDFs/slides needed refined post-processing in the Lambda code.
  
• Terraform & IAM pain — Getting least-privilege permissions right for Bedrock, S3, and DynamoDB involved a lot of trial-and-error; one missing action blocked the entire flow for hours.
  
• Time pressure — With the March 16, 2026 deadline looming, I had to ruthlessly prioritize: focus on a rock-solid core loop (upload → process → summary/quiz/chat) and skip nice-to-haves to ship a complete, demo-ready MVP.

Accomplishments that I'm proud of

• 100% serverless architecture deployed cleanly via Terraform — reproducible and zero server management
  
• Effectively using Nova 2 Lite’s 1M token context to keep full lecture content available during chat without any chunking or external RAG
  
• Clean, modern dark-mode UI that feels intuitive and student-friendly (inspired by Notion + ChatGPT)
  
• Secure flow end-to-end: direct browser-to-S3 uploads, private per-user data, Cognito-protected APIs
  
• Smooth 3-minute demo: upload lecture → instant Nova processing → quiz + intelligent chat

What I learned

• Nova 2 Lite is surprisingly powerful for multimodal document reasoning at a very low cost — its large context and vision capabilities make it ideal for education use cases without needing massive models.
  
• Serverless requires careful design (presigned uploads, timeout tuning, structured outputs) to feel fast and seamless to users.
  
• The quality of AI output depends heavily on prompt quality — small changes in wording, structure, and examples make a huge difference.
  
• Hackathons are intense but teach you to ship fast, debug under pressure, and focus on what actually matters for the demo and judges.

What's next for NovaBuddy

• Improve quiz variety and difficulty levels (adaptive based on user performance)
  
• Add semantic search across multiple lectures using Nova Multimodal Embeddings
  
• Export summaries/quizzes as clean PDFs or Anki flashcards
  
• Allow sharing read-only versions of processed lectures for study groups
  
• Write and publish a builder.aws.com blog post about how tools like NovaBuddy can help students in under-resourced educational environments → hopefully qualify for the bonus AWS credits prize
  
• Open-source the repository and gather feedback from fellow students to keep improving

Huge thanks to the Amazon Nova team — this was my first deep dive into frontier multimodal models, and NovaBuddy simply wouldn’t exist without them.

Built With

• cognito
• lambda
• terraform