MultiAgent GenAI Tutor

Live Demo - Deployed Link

Video Demonstration of MultiAgent GenAI Tutor

A Flask-powered web application that implements a multi-agent tutoring assistant using Google’s Gemini LLM and specialized tools. Users can ask questions in math, physics, chemistry, or biology; upload images for OCR; or paste YouTube URLs for AI-generated summaries.

---

Table of Contents

1. Features
2. Architecture & Flow
3. Setup & Installation
4. Environment Variables
5. Running Locally
6. Deployment
7. Folder Structure
8. Screenshots
9. Agent–Tool Interaction
10. Challenges & Solutions

---

Features

• Text Chat: Ask any question in math, physics, chemistry, or biology.
• OCR: Upload an image of a question; gets converted to text.
• YouTube Summarizer: Paste a youtube.com or youtu.be link and receive concise bullet points.
• Calculator: Automatic math expression evaluation via CalculatorTool.
• Email Transcript: Receive a PDF copy of your full chat via email.

---

Architecture & Flow

1. User Input

   • Text question → MentorAgent
   • Image upload → OCRTool → text → MentorAgent
   • YouTube URL → YouTubeSummarizerTool

2. MentorAgent

   • Detects YouTube URLs (regex).
   • Otherwise, classifies the subject (math/physics/chemistry/biology) via a Gemini prompt.

3. TutorAgent

   • Routes by subject to the appropriate agent (MathAgent, PhysicsAgent, BiologyAgent).
   • If subject is unknown or unsupported, returns a friendly fallback.

4. Subject Agents

   • MathAgent: Checks for numeric expressions, uses CalculatorTool if found; else asks Gemini.
   • PhysicsAgent/BiologyAgent/ChemistryAgent: Delegate directly to Gemini with subject-specific prompts.

5. Tools

   • CalculatorTool: Safe eval of math expressions.
   • OCRTool: PIL + Tesseract OCR.
   • YouTubeSummarizerTool: Fetches transcript, uses Gemini to summarize.
   • EmailTool: Converts chat to PDF and emails it.

---

Setup & Installation

1. Clone

   git clone https://github.com/yourusername/MultiAgent-Tutor-GenAI.git
   cd MultiAgent-Tutor-GenAI

2. Environment

   python3 -m venv venv
   source venv/bin/activate   # Windows: venv\Scripts\activate
   pip install --upgrade pip
   pip install -r requirements.txt
   
   Tesseract OCR
   Ubuntu: sudo apt-get install tesseract-ocr
   Windows: Download MSI from https://github.com/tesseract-ocr/tesseract

Tesseract OCR

Ubuntu

sudo apt-get install tesseract-ocr

Windows

Download and install from: https://github.com/tesseract-ocr/tesseract

---

Environment Variables

Create a .env file in the project root:

# Gemini LLM API
GEMINI_API_KEY=your_google_gemini_api_key

# Flask session
FLASK_SECRET_KEY=a_random_secret_key

# (Optional) EmailTool SMTP
SMTP_SERVER=smtp.gmail.com
SMTP_PORT=587
SENDER_EMAIL=your_email@example.com
SENDER_PASSWORD=your_email_password

---

Running Locally

export FLASK_APP=app.py
export FLASK_ENV=development
flask run

Visit http://localhost:5000 in your browser.

---

Deployment

Deployed on Render. Live App: https://multiagent-tutor-genai-1.onrender.com/

---

Folder Structure

MultiAgent-Tutor-GenAI/
├── agents/
│   ├── mentor_agent.py      # Orchestrator
│   ├── subject_classifier.py
│   ├── tutor_agent.py
│   ├── math_agent.py
│   ├── physics_agent.py
│   ├── biology_agent.py
│   └── chemistry_agent.py
├── tools/
│   ├── calculator_tool.py
│   ├── ocr_tool.py
│   ├── youtube_summarizer.py
│   └── email_tool.py
├── templates/
│   └── index.html
├── static/
│   └── ss1.png
│   └── ss2.png
├── .env
├── requirements.txt
└── app.py

---

Screenshots

Agent–Tool Interaction

Flow:

User → MentorAgent → [OCRTool | Subject Classifier] → TutorAgent → [CalculatorTool or Gemini LLM] → Response

Extensibility:

Add new agents (e.g. HistoryAgent) by implementing answer() and registering it in tutor_agent.py.

Prompt Control:

Carefully engineered prompts ensure consistent outputs:

• Single-word classification
• 250-word bullet summaries

---

Challenges & Solutions

Subject Classification Errors

Fix: Prompt refinement to enforce exact category keywords and handle “unknown” fallback.

OCR Misreads

Fix: Pre-processing images (contrast adjustments) and error handling for blank extractions.

LLM Hallucinations in Math

Fix: Offload pure computations to the CalculatorTool whenever possible.

Session Size

Fix: Consider truncating very long transcripts or migrating to a database for persistence.

---