██╗ ███████╗ █████╗ ██████╗ ███╗ ██╗███████╗ ██████╗ ██████╗ ██████╗ ███████╗
██║ ██╔════╝██╔══██╗██╔══██╗████╗ ██║██╔════╝██╔═══██╗██╔══██╗██╔════╝ ██╔════╝
██║ █████╗ ███████║██████╔╝██╔██╗ ██║█████╗ ██║ ██║██████╔╝██║ ███╗█████╗
██║ ██╔══╝ ██╔══██║██╔══██╗██║╚██╗██║██╔══╝ ██║ ██║██╔══██╗██║ ██║██╔══╝
███████╗███████╗██║ ██║██║ ██║██║ ╚████║██║ ╚██████╔╝██║ ██║╚██████╔╝███████╗
╚══════╝╚══════╝╚═╝ ╚═╝╚═╝ ╚═╝╚═╝ ╚═══╝╚═╝ ╚═════╝ ╚═╝ ╚═╝ ╚═════╝ ╚══════╝
Built for the Google × Kaggle 5-Day AI Agents Intensive Capstone 2026
LearnForge is an AI study assistant that helps software engineering students learn smarter. You pick your engineering track — CSE, IT, AI/ML, or Data Science — and four AI agents work together to tutor you, quiz you, plan your study schedule, and review your code.
The entire system runs on your own computer using Gemma 3 (a free local AI model). No internet required after setup. No API keys. No monthly fees.
Think of it as having a personal AI tutor available 24/7 that knows exactly what you need to study next.
| Student type | What LearnForge helps with |
|---|---|
| CSE students | Data Structures, Algorithms, OS, DBMS, Networks, System Design |
| IT students | Networking, Cybersecurity, Cloud Computing, DevOps, Linux |
| AI/ML students | ML Theory, Deep Learning, NLP, LLMs, MLOps |
| Data Science students | Statistics, Pandas, SQL, EDA, Feature Engineering |
You type a question or request
↓
Orchestrator Agent figures out what you need
↓
Sends your request to the right specialist agent
↓
┌─────────────┬──────────────┬──────────────┬───────────────┐
│ Tutor Agent │ Quiz Agent │Planner Agent │Code Review │
│ │ │ │Agent │
│ Explains │ Gives you │ Builds your │ Reads your │
│ concepts │ MCQ tests │ weekly study │ code and │
│ clearly │ and tracks │ plan │ fixes it │
│ │ weak areas │ │ │
└─────────────┴──────────────┴──────────────┴───────────────┘
↓
Agents use tools (search papers, run your code, check your progress)
↓
Answer comes back to your browser
This project was built for the Google × Kaggle Vibe Coding Capstone 2026. Here is exactly where each requirement is demonstrated:
| # | Requirement | What was built | Where to find it |
|---|---|---|---|
| 1 | Multi-agent system (ADK) | Orchestrator + 4 specialist agents built with Google ADK | backend/agents/ |
| 2 | MCP Server | Local Python MCP server with 5 tools agents can call | backend/mcp/server.py |
| 3 | Antigravity | Entire project generated using Antigravity CLI agent | Video demo |
| 4 | Security features | JWT auth, prompt injection guard, sandboxed code execution, rate limiting | backend/security/ |
| 5 | Deployability | Railway.app auto-deploy + Docker + one-command local setup | railway.toml, Dockerfile |
| 6 | Agent skills (CLI) | learnforge-cli with 4 working commands |
cli/learnforge_cli.py |
This is the "brain" that receives your message, decides which specialist to send it to, and coordinates everything.
You: "Quiz me on transformers, I'm intermediate level"
↓
Orchestrator reads your message
↓
Decides → this needs the Quiz Agent
↓
Passes: topic=transformers, level=intermediate
↓
Quiz Agent takes over
Explains any concept in a way that actually makes sense. Always gives:
- A simple 2-sentence intuition first
- Key bullet points
- A worked example or code snippet
- One real-world application
Generates 4-option MCQs adapted to your level. Tracks which topics you keep getting wrong and focuses on those next time (spaced repetition).
Builds you a personalised 7-day or 30-day study plan based on your track and your past quiz performance. If you're weak on a topic, it schedules more time there automatically.
Paste your Python, JavaScript, or SQL code. The agent runs it in a safe sandbox, checks for bugs, style issues, and time complexity, then gives you a fixed version.
The MCP (Model Context Protocol) server is a local service that gives agents access to real tools, not just the language model.
MCP Server running locally on port 8001
│
├── search_arxiv(query)
│ → Searches academic papers so the Tutor Agent can cite real research
│
├── run_code_sandbox(code)
│ → Safely runs your code in an isolated box with a 10-second timeout
│
├── get_student_progress(student_id)
│ → Reads your quiz history from the local database
│
├── generate_quiz(topic, level, weak_areas)
│ → Asks Gemma 3 to create fresh MCQs tailored to your weak spots
│
└── get_concept_graph(topic, track)
→ Returns what topics you should learn before tackling this one
Your message comes in
↓
① Rate limiting — max 20 messages/min so no one can spam the system
↓
② JWT authentication — you must be logged in with a valid token
↓
③ Prompt injection guard — strips attempts to hijack the AI with tricks
Example: "ignore previous instructions" gets caught and removed
↓
④ Sandboxed code execution — your code runs in a locked box:
- 10 second timeout (kills infinite loops)
- 64MB RAM limit
- No internet access from inside the sandbox
- No file system access outside /tmp/sandbox
↓
Safe response sent back to you
learnforge-AI/
│
├── backend/ ← Python server (FastAPI)
│ ├── agents/
│ │ ├── orchestrator.py ← Routes your request to the right agent
│ │ ├── tutor_agent.py ← Explains concepts
│ │ ├── quiz_agent.py ← Generates and grades MCQs
│ │ ├── planner_agent.py ← Builds study plans
│ │ └── code_review_agent.py← Reviews and fixes code
│ │
│ ├── mcp/
│ │ ├── server.py ← MCP server (gives agents tools to use)
│ │ └── tools/
│ │ ├── arxiv_tool.py ← Paper search
│ │ ├── sandbox_tool.py ← Safe code execution
│ │ ├── progress_tool.py← Read/write your progress DB
│ │ └── quiz_tool.py ← MCQ generation via Gemma 3
│ │
│ ├── security/
│ │ ├── auth.py ← Login tokens (JWT)
│ │ ├── sanitizer.py ← Blocks prompt injection attacks
│ │ └── rate_limiter.py ← Limits requests per minute
│ │
│ ├── db/
│ │ └── models.py ← Database table definitions
│ │
│ ├── api/
│ │ └── routes.py ← All API endpoints (/chat, /quiz, etc.)
│ │
│ ├── main.py ← Starts the FastAPI server
│ ├── requirements.txt ← Python packages needed
│ └── Dockerfile ← How to package the backend
│
├── frontend/ ← React web app (what you see in browser)
│ └── src/
│ ├── components/
│ │ ├── HomeDashboard.jsx ← Main page with project info
│ │ ├── TrackSelector.jsx ← Pick CSE / IT / AI-ML / DS
│ │ ├── TutorChat.jsx ← Chat with the AI tutor
│ │ ├── QuizMode.jsx ← Take adaptive MCQ tests
│ │ ├── StudyPlanner.jsx ← See your weekly study plan
│ │ ├── CodeReview.jsx ← Paste code, get feedback
│ │ └── Analytics.jsx ← See your progress charts
│ └── App.jsx ← Main app file
│
├── cli/
│ └── learnforge_cli.py ← Command line tool (Agent skill)
│
├── .github/workflows/
│ └── deploy.yml ← Auto-deploy to Railway on every push
│
├── docker-compose.yml ← Run everything locally in one command
├── railway.toml ← Railway deployment config
├── .env.example ← Template for your environment variables
└── README.md ← This file
- Python 3.11+ — download from https://python.org
- Node.js 18+ — download from https://nodejs.org
- Ollama — download from https://ollama.ai (this runs the AI model)
- Git — download from https://git-scm.com
git clone https://github.com/Butkii025/learnforge-AI.git
cd learnforge-AIcp .env.example .envOpen .env and fill it in:
SECRET_KEY=make-up-any-long-random-string-here
OLLAMA_HOST=http://localhost:11434
DATABASE_URL=sqlite:///./db/progress.db
CORS_ORIGINS=http://localhost:3000ollama pull gemma3
ollama serveLeave this terminal running. Open a new one for the next steps.
cd backend
python -m venv venv
venv\Scripts\activate # Windows
# source venv/bin/activate # Mac/Linux
pip install -r requirements.txt
uvicorn main:app --reload --port 8000Open another terminal:
cd backend
venv\Scripts\activate
python mcp/server.pyOpen another terminal:
cd frontend
npm install
npm run devGo to http://localhost:3000 — LearnForge is running!
Instead of all the above steps, just run:
docker compose upThen open http://localhost:3000
pip install -e ./cliThen you can run:
# Get a topic explained
learnforge study --topic "transformer attention mechanism" --track aiml
# Take a quiz
learnforge quiz --topic "backpropagation" --difficulty hard
# Generate a study plan
learnforge plan --goal "master MLOps in 30 days" --track aiml
# Get code reviewed
learnforge review --file my_model.py
***
***
***
Video.Project.6.mp4
| Part | Technology | Why this was chosen |
|---|---|---|
| AI model | Gemma 3 via Ollama | Runs 100% locally, no internet or API key needed |
| Agent framework | Google ADK | Competition requirement, clean multi-agent pattern |
| Tool protocol | MCP Python SDK | Competition requirement, lets agents share tools |
| Backend | FastAPI + Python 3.11 | Fast, modern, great for streaming AI responses |
| Frontend | React 18 + Vite | Fast builds, great developer experience |
| Styling | Tailwind CSS | No CSS files to maintain |
| Database | SQLite + SQLAlchemy | Zero config, stores progress locally |
| Authentication | JWT (python-jose) | Secure, stateless, works on Railway |
| Deployment | Railway.app | Free tier, GitHub auto-deploy |
| CI/CD | GitHub Actions | Tests run automatically on every push |
| CLI | Python + argparse | Agent skills competition requirement |
Most AI study tools are just a chatbot wrapper around GPT. LearnForge is different in three ways:
1. It knows your discipline. When you pick AI/ML as your track, every topic, quiz, study plan, and tutor response is calibrated for AI/ML. A CSE student gets completely different content.
2. It actually remembers your progress. Your quiz results are stored locally. The Quiz Agent reads them before generating questions, so it focuses on what you actually struggle with.
3. It runs your code for real. The Code Review Agent doesn't just read your code — it executes it in a sandbox and shows you the actual output alongside the feedback.
- Google × Kaggle 5-Day AI Agents Intensive Course 2026
- Google ADK team for the multi-agent framework
- Ollama project for making local LLMs accessible
- MCP SDK contributors
Priynashu Vijay — Built for Google × Kaggle Vibe Coding Capstone 2026
