Roast2Learn

"The AI tutor that roasts you into a genius."

Comment 1

💡 Inspiration

Most study apps feel like corporate productivity software. They're powerful but soulless — and students forget what they study with them.

The science is clear: memory is emotional. We remember things tied to embarrassment, humor, or surprise — not flashcards we passively flipped through. So we built the opposite of every study app on the market: an AI tutor that personality-tests every wrong answer with a witty roast, then teaches the right answer.

This isn't gimmicky. Cognitive science calls it emotional encoding — and it's why we remember our most embarrassing moments perfectly while forgetting last Tuesday's lecture.

🎯 What it does

Roast2Learn turns any messy PDF, photo of handwritten notes, or slide deck into a complete study session with 10 AI-powered tools, organized by cognitive learning phase:

🧠 Understand

  • AI Summary — TLDR, key concepts, must-remember facts, common pitfalls
  • Smart Notes Rewriter — Messy notes turned into clean structured markdown

📚 Absorb

  • Live AI Tutor — Chat about your uploaded material. Get roasted when wrong.
  • Auto Flashcards — 10-50 cards per set, optimized for active recall

🔥 Test

  • Quiz + Roast Mode — Wrong answer = personalized AI roast + explanation
  • Step-by-Step Solver — Upload any problem (image, PDF, text), get every step worked out

🎯 Diagnose

  • Weak Spot Diagnosis — AI analyzes your history, identifies what you actually don't know

🔮 Predict

  • Prof-Style Questions — Upload a past exam, AI generates questions in your professor's tone
  • Exam Topic Predictor — Past papers turn into ranked predictions for what's next

📋 Finalize

  • 1-Page Cheat Sheet — Distills a semester of notes into one dense, exam-ready page

🔥 NEW: Study Streak (Built During Hackathon)

  • Tracks consecutive days of AI tool usage
  • Duolingo-style streak mechanics drive daily return engagement
  • Visible badge on the workspace home with current streak and status

🛠️ How we built it

Stack: Next.js 14 (App Router) · Supabase (auth + DB + storage) · OpenAI GPT-4o-mini + Vision · Stripe · Vercel

Architecture highlights:

  • Server-side rendering for instant page loads
  • Per-action rate limiting with daily quotas tracked live from DB counters
  • Single source of truth for tier resolution (Stripe subscription > active promo > free)
  • Sanitization layer for Unicode-heavy JSONB writes (handwriting/equation parsing)
  • Cached AI generations to control cost
  • Edge-deployed image vision for handwritten/photo notes

Smart cost engineering:

  • Each AI tool caches results so re-opening doesn't burn tokens
  • Free tier hard-capped per action per day, resetting at midnight
  • Daily limits chosen over monthly to drive return visits

🚧 Challenges we ran into

Personality without cruelty. The hardest challenge was making the AI feel like a real personality without being mean to users. Early versions had the AI insulting the student instead of the wrong answer — that crossed a line. We rewrote the system prompts multiple times before landing on the current tone: targets the answer, never the person, always educational, always ends by teaching the correct concept.

Rate limit architecture. We initially built a single "AI usage" counter, but realized different tools have different cost profiles. Vision-based problem solving costs significantly more than flashcard generation. So we refactored to per-action daily limits, with each tool's quota tracked independently from live database row counts — meaning limits are always accurate and can't drift.

Stripe webhook reliability. Subscription state needs to be the single source of truth for tier, but webhooks can fire out of order or be delayed. We built computeEffectiveTier() as a deterministic resolution function: reads Stripe subscription status first, then active promo codes, then defaults to free. Tier never drifts, even with webhook race conditions.

🏆 Accomplishments we're proud of

Already monetizing. Stripe is live with real transactions verified end-to-end. Free → Pro → Beast tier transitions work via webhook with zero manual intervention.

10 fully-working AI tools. Most hackathon projects have 1-2 demos. Roast2Learn has Summary, Smart Notes, Live Tutor, Flashcards, Quiz+Roast, Solver, Weak Spot Diagnosis, Prof-Style Questions, Exam Predictor, and Cheat Sheet — every one tested in production.

Production-quality polish. Custom gradient brand, dark mode, mobile responsive, custom favicon, branded auth emails, working password reset flow, account deletion. Not hackathon-tier prototype quality.

A genuinely original idea. We searched extensively for similar products with personality-first AI in studying. Nothing exists at this level of execution. Roast2Learn fills a real gap.

🎓 What we learned

Personality is a feature, not a bug. We worried at first that "an AI that roasts you" would feel unprofessional. The opposite happened: every test user laughed at their roasts and remembered the lesson better. Personality made the product 10x more engaging and memorable.

Cached generations are everything. OpenAI calls aren't free. Every AI tool caches its result so re-opening a generated summary doesn't burn another API call. This made the unit economics work at $8.99/month.

Webhook architecture matters. Stripe webhooks need to be idempotent and order-independent. We learned to never trust webhook order and always resolve tier from current database state, not from what the webhook said happened.

🚀 What's next for Roast2Learn

  • 🎤 Voice Tutor Mode — study hands-free with speech-to-text and AI voice
  • 🎙️ Live lecture recording — real-time notes and flashcards
  • 📱 Mobile apps (iOS and Android)
  • 👥 Study rooms — shared quizzes between friends
  • 🧠 Spaced-repetition scheduling
  • 🌍 Multi-language support (Arabic, Spanish, French)
  • 🏫 University pilot program

🚧 HACKATHON DISCLOSURE

Full transparency, per Techspire rules ("Previously built projects are not allowed unless disclosed"):

Roast2Learn was started as a personal side project before the hackathon period. The base platform — 10 AI tools, authentication, Stripe payments, and core infrastructure — pre-existed.

The following was built specifically during the hackathon period (June 1, 2026):

  1. 🔥 Study Streak feature — a new retention feature using Duolingo-style daily streak mechanics, tracking consecutive days of AI tool usage with persistence in the user profile schema
  2. 📚 Comprehensive technical documentation
  3. 🎥 Demo video and presentation materials
  4. 🐛 Final round of polish and bug fixes

The pre-existing base demonstrates a complete production system; the hackathon-period work shows continued innovation and execution.

🤖 AI Tools Used During Development

Per disclosure norms in modern hackathon competitions:

Claude (Anthropic) was used by our team for:

  • Code review and architecture guidance
  • Documentation drafting
  • Pitch material preparation
  • Debugging assistance

All code was written, tested, and deployed by our team members. We used Claude as a coding assistant the way modern engineering teams use AI tools.

OpenAI GPT-4o and Vision APIs power the in-app features users interact with.

📦 Open Source Attribution

Roast2Learn is built on the shoulders of incredible open-source work:

  • Next.js (Vercel) — React framework
  • Supabase — Open-source Firebase alternative for auth, DB, storage
  • PostgreSQL — Relational database
  • React — UI library
  • Stripe — Payment processing
  • html2canvas — Used for shareable card generation

All used per their respective licenses.

👥 Team

Built by Syed Affan Hussaini and Azmi Arafat Sultan — students who got tired of using study apps that felt like spreadsheets.

Built With

Share this project:

Updates