Glitch

let's break the system

An adaptive learning platform that rewires any PDF around the way YOUR brain works.

---

What is Glitch?

Glitch is an adaptive learning platform powered by GBrain that builds a personalized curriculum from any PDF and teaches you in a way that matches how YOUR brain works.

Upload a textbook. Tell Byte what you're into. Watch your interests transform the content.

🎮 Pokémon fan studying immunology? Your macrophages become Pokémon defenders. 🏎️ F1 fan learning AI? Your neural networks race like lap times. ⚡ Percy Jackson fan tackling calculus? Limits become demigod quests.

---

✨ The Magic

• 🧬 Same PDF, completely different experience — Alex (visual, beginner, loves Pokémon) and Sam (theory, advanced, loves F1) get entirely different curricula, lesson styles, and content framing from identical source material.
• 🧠 GBrain-powered memory — every checkpoint answer is saved back to GBrain. Get something wrong, come back later, Byte remembers and rebuilds the lesson specifically around your weak spots.
• 🎨 AI-generated diagrams — visual learners get real AI-generated illustrations for every lesson section.
• 🤖 Personalized loading commentary — Byte cracks jokes connecting your interests to the topic while lessons load. No two waiting screens are the same.
• 📚 Multi-course isolation — each uploaded course lives in its own GBrain source namespace. No bleed-over between immunology and machine learning.
• 🔁 Adaptive review mode — return to a topic and Byte knows what you got wrong last time. The next lesson is rebuilt around those exact concepts.

---

🔧 How It Works

flowchart TD
    A[📄 PDF Upload] -->|pymupdf4llm| B[Markdown Conversion]
    B -->|gbrain import --source-id| C[(GBrain<br/>Course Namespace)]
    D[👤 Student Profile<br/>style · level · interests] --> E[Curriculum Generator]
    C -->|hybrid search<br/>vector + keyword + RRF| E
    E -->|Claude Sonnet| F[Personalized Curriculum]
    F --> G[Topic Selected]
    G --> H[Lesson Generator]
    C --> H
    I[(GBrain<br/>Student Progress)] -->|past wrong answers| H
    H -->|Claude Sonnet| J[Adaptive Lesson<br/>3 checkpoints]
    J -->|visual learners only| K[OpenAI gpt-image-2<br/>diagrams]
    J --> L[Student answers]
    L -->|gbrain put| I
    L -->|trigger review mode| H

Loading

Step-by-step

1. PDF upload → pymupdf4llm converts the deck to clean markdown → GBrain ingests it with --source-id <slug> for hard isolation between courses.
2. GBrain hybrid search (vector embedding + keyword + Reciprocal Rank Fusion) retrieves relevant chunks per query, scoped to a single course.
3. Curriculum generation — student profile (learning style, level, interests) + retrieved course content → Claude produces a JSON curriculum ordered for that specific learner.
4. Each lesson — GBrain retrieves topic-specific content plus the student's past progress page → Claude generates sections, examples, and 3 checkpoints, each testing a different concept.
5. Checkpoint answers — saved back to GBrain as a markdown page at students/<slug>-progress, tagged with the conceptTested field for later retrieval.
6. Review mode — when you return to a topic, the client passes previousAttempts from localStorage; the API switches to a review prompt that focuses the new lesson on whatever you got wrong.
7. Visual learners — Claude's lesson output is run through OpenAI's gpt-image-2 (via the Responses API) to generate clean educational diagrams for the first two sections.
8. Loading screens — Claude Haiku generates a batch of 10 personalized punny messages connecting your interests to the topic, with a silent background refill before the queue runs out.

---

🧪 Tech Stack

Layer	Tech
Framework	Next.js 14 (App Router, React Server Components)
Language	TypeScript
Styling	Tailwind CSS + custom CSS animations
Knowledge layer	GBrain — persistent knowledge graph, hybrid RAG, student memory
Curriculum & lessons	Claude (claude-sonnet-4-20250514) via Anthropic SDK
Loading vibes	Claude Haiku (claude-haiku-4-5-20251001) — fast, cheap, punny
Diagrams	OpenAI (gpt-image-2 via the Responses API)
PDF parsing	pymupdf4llm
Profile storage	localStorage (no server-side user DB)

---

🔑 API Keys Required

Create a .env.local in the project root:

ANTHROPIC_API_KEY=sk-ant-...     # https://console.anthropic.com
OPENAI_API_KEY=sk-proj-...       # https://platform.openai.com
GEMINI_API_KEY=AIza...           # https://aistudio.google.com (optional fallback)

---

🚀 Setup

# 1. Install dependencies
npm install

# 2. Install GBrain
git clone https://github.com/garrytan/gbrain
cd gbrain && bun install && bun link
cd ..

# 3. Initialize GBrain with PGLite (local, no external DB)
gbrain init --pglite

# 4. Install the PDF parser
pip install pymupdf4llm

# 5. Add API keys to .env.local (see above)

# 6. Run the dev server
npm run dev

Then visit localhost:3000 → onboard → upload a PDF → start learning.

---

🎭 Demo Personas

Three pre-built profiles to feel the difference in seconds. Pick one on the onboarding screen.

Persona	Learning Style	Level	Lesson Length	Interests	What you'll see
Alex 🎴	Visual	Beginner	Short	Pokémon	Emoji-rich lessons, AI-generated diagrams, gentle pacing, Pokémon-themed analogies
Sam 🏁	Theory	Advanced	Deep dives	F1 racing	Dense academic prose, no images, technical depth, racing metaphors
Jordan ⚡	Examples	Intermediate	Medium	Percy Jackson	Narrative-driven explanations, mythology-flavored examples

Try this: upload the same PDF, switch between Alex and Sam, generate a curriculum twice. The two experiences barely overlap.

---

📁 Project Structure

glitch/
├── app/
│   ├── page.tsx                  # Landing — Glitch hero + Byte
│   ├── onboarding/page.tsx       # 5-question chat with Byte
│   ├── dashboard/page.tsx        # Course selector + personalized curriculum
│   ├── upload/page.tsx           # PDF ingest with vibe commentary
│   ├── lesson/page.tsx           # Adaptive lesson + 3 checkpoints
│   └── api/
│       ├── curriculum/route.ts   # GBrain → Claude → JSON curriculum
│       ├── lesson/route.ts       # Lesson gen, answer eval, image gen
│       ├── checkpoint/route.ts   # Persist checkpoint results to GBrain
│       ├── ingest/route.ts       # PDF → markdown → GBrain import
│       ├── vibe/route.ts         # Claude Haiku loading commentary
│       └── chat/route.ts         # General Byte chat endpoint
├── components/
│   ├── Byte.tsx                  # Mascot — 4 mood-keyed PNGs
│   ├── SpeechBubble.tsx
│   └── PillButton.tsx
└── lib/
    ├── gbrain.ts                 # GBrain CLI wrapper + source-id filter
    ├── claude.ts                 # Anthropic SDK + JSON extractor
    └── profile.ts                # localStorage helpers, slugify, types

---

🏗️ Built at YC GStack × GBrain Hackathon

Built in one day using GStack for the development workflow and GBrain as the core memory and RAG layer. The whole point of the demo: prove that personalized, adaptive learning is a thing you can ship in 24 hours when the retrieval and memory layers are already solved for you.

---

Built by humans, taught by a tiny gremlin named Byte.