Automated Marking System for Algebraic Math Solutions

An AI-powered system that automatically grades multi-step handwritten math solutions with step-by-step feedback. Combines OCR, symbolic reasoning, and a fine-tuned LLM to evaluate algebraic problems like a human instructor.

---

🏆 Project Recognition

Best Major Project Award - IOE Thapathali Campus (2025)

Recognized as one of the best major projects among 12 project in BCT batch 078

---

🎯 Project Showcase

System Interface

📹 Demo Video

Video.Project.4.mp4

---

📋 Table of Contents

• Project Showcase
• Project Recognition
• What It Does
• Key Features
• Dataset
• System Architecture
• Tech Stack
• Training
• Results Showcase
• Quick Start
• Documentation
• Team

What It Does

• 📸 Reads handwritten math → Converts to digital format using OCR
• ✓ Validates each step → Checks mathematical correctness symbolically
• 🐛 Detects errors → Identifies arithmetic mistakes and invalid transformations
• 📊 Awards partial marks → Gives credit for correct steps, even if final answer is wrong
• 💬 Generates feedback → Explains why answers are right or wrong

Key Features

✨ Step-by-step validation using symbolic reasoning (SymPy)
✨ Partial marking algorithm based on solution quality
✨ Error propagation detection
✨ Mathematically equivalent answer recognition
✨ Natural language feedback generation
✨ Web-based interface for easy image upload
✨ Supports algebraic simplification, equations, and logarithms

Dataset

Total Samples: 2,797 question-rubric pairs (after augmentation)

Distribution by Topic

*Bar chart showing distribution across Simplification, Equations, and Logarithms*

Topic	Samples	%	Difficulty
Algebraic Simplification (SIM)	735	26%	Easy, Medium, Hard
Algebraic Equations (EQ)	1286	46%	Easy, Medium, Hard
Logarithmic Expressions (LOG)	776	28%	Easy, Medium, Hard

Distribution by Difficulty Level

Distribution across difficulty levels within each topic

• Easy (1 mark): ~33% of samples
• Medium (2 marks): ~33% of samples
• Hard (>2 marks): ~34% of samples

Data Sources

Source	Samples	Type
NEB Official Exam Papers	89	Real exams (2080-2082)
Curriculum Development Center (CDC)	45	Official guidelines
Class 7-10 Textbooks	78	Educational materials
Reference Problem Sets	67	Competitive math resources
Original Collected	219	Manually verified
Augmented via
- Variable substitution	+450	x, y, a, b → different vars
- Coefficient scaling	+380	Numeric manipulation
- Structural variations	+340	Reordered polynomials
- Factorization templates	+468	Standard patterns
- Exponent law templates	+340	Power rules

Data Pipeline: Collection → OCR Extraction → Manual Verification → Cleaning → Augmentation → 2,797 final samples

System Architecture

Pipeline Diagram

*Complete system architecture showing data flow from input image to final output*

What to show here:

• Block diagram from your Project Report (Figure 4-1)
• Shows: Image → OCR → Parser → Validation → LLM → Output
• Visual representation of how all components connect

Component Overview

Handwritten Image
      ↓
  [Mathpix OCR] ──→ LaTeX Format
      ↓
  [Regex Preprocessing] ──→ Clean Steps
      ↓
  [SymPy Parser] ──→ Symbolic Objects
      ↓
  [Symbolic Reasoning Engine] ──→ Step Validation
      ↓
  [DeepSeekMath-7B] ──→ Rubric Generation
      ↓
  [LLM Marking Engine] ──→ Score + Feedback

Key Components:

• Symbolic Engine: Ensures mathematical correctness
• LLM: Generates human-like explanations & rubrics
• Partial Marking: Rewards partial progress algorithmically

Tech Stack

Layer	Technology
Backend	Python, Flask, SymPy, DeepSeekMath-7B
Frontend	React, Tailwind CSS
ML/AI	PyTorch, Hugging Face, LoRA fine-tuning
OCR	Mathpix API
Infrastructure	Google Colab

Training

Model Training Curves

Training Configuration:

• Model: DeepSeekMath-7B with LoRA (rank 32)
• Dataset: 2,237 training samples (80%)
• Epochs: 5 with early stopping
• Optimizer: AdamW (lr=0.0005)
• Batch Size: 32

Training and validation loss convergence over 5 epochs

Performance Metrics

Metric	Score	Details
OCR Accuracy	95%	Character error rate on clear handwriting
Symbolic Validation	100%	Step-by-step mathematical correctness
Partial Marking Accuracy	90%	Alignment with human rubrics
Parse Success Rate	98%	Valid JSON output generation
Model Convergence	✓	Stable training with minimal overfitting

Results Showcase

These are actual output examples from the system when grading different solutions:

Example 1: Perfect Solution ✅

Perfect solution (3/3 marks) - All steps correct, full feedback provided

What's shown: Input math image → Output: Full marks, positive feedback, congratulatory message

---

Example 2: Partial Credit ⚠️

Partial credit (3.25/4 marks) - Some steps correct, some wrong, detailed error analysis

What's shown: Input math image → Output: Partial marks, error highlighting, which step went wrong

---

Example 3: Error Detection ❌

Error detection (0/2 marks) - Critical error caught, learning feedback provided

What's shown: Input math image → Output: zero marks, explanation of where calculation failed, suggestions

---

Key Achievements:

• ✅ Detects arithmetic errors in intermediate steps
• ✅ Recognizes mathematically equivalent forms
• ✅ Provides step-specific feedback and error propagation analysis
• ✅ Maintains consistency in grading

Quick Start

Prerequisites

• Python 3.8+
• Node.js 14+
• Mathpix API key (get one here)

Installation

# Clone repository
git clone https://github.com/yourusername/math-autograder.git
cd math-autograder

# Backend setup
python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate
pip install -r requirements.txt

# Set environment variables
cp .env.example .env
# Edit .env with your Mathpix credentials

# Run backend
flask --app run run

# In another terminal: Frontend setup
cd frontend
npm install
npm start

Open http://localhost:3000 → Upload image → Get instant marks + feedback

Documentation

📄 Full Technical Details: See Project_Report.pdf

• Complete methodology & algorithm details
• Dataset construction (2,797 NEB exam samples)
• Model fine-tuning process (LoRA)
• Comprehensive evaluation metrics
• Implementation code snippets
• Detailed results analysis
• Future improvements roadmap

Team

Name	Roll
Prajwal Chaudary	THA078BCT028
Jesis Upadhayaya	THA078BCT017
Purushottam Gajurel	THA078BCT032
Sagar Bikram Adhikari	THA078BCT037

Supervisor: Asst. Prof. Suwarna Lingden
Institution: Institute of Engineering, Tribhuvan University (2025)

License

MIT License - see LICENSE file

⭐ If this project helped you, please star it on GitHub!