💰 TransactAI: Personal Expense Categorization System

AI-Powered Bank Statement Analysis & Expense Categorization

TransactAI is an intelligent expense tracking system that automatically categorizes bank transactions using a hybrid approach combining DistilBERT deep learning with context-aware NLP rules. Built as an M.Sc Data Science project, it achieves 95%+ accuracy on real Indian bank statements.

---

📋 Table of Contents

• Features
• Demo
• How It Works
• Installation
• Usage
• Project Structure
• Model Architecture
• Results
• Technologies Used
• Contributing
• License
• Contact

---

✨ Features

• 🤖 AI-Powered Categorization: Fine-tuned DistilBERT model combined with rule-based NLP for high accuracy
• 📊 16+ Categories: Food, Shopping, Travel, Bills, Income, Cashback, EMI, Utilities, Healthcare, and more
• 🏦 Multi-Bank Support: Works with HDFC, ICICI, SBI, Federal Bank, Axis Bank, and other Indian banks
• 📁 Multiple Formats: Supports CSV and Excel (.xlsx, .xls) statements
• 📈 Interactive Analytics: Real-time pie charts, bar charts, spending insights, and personalized recommendations
• 📄 Professional PDF Reports: Generate comprehensive expense analysis reports with charts and insights
• 🔄 Smart Parsing: Automatically detects headers and handles complex Excel formats with metadata
• 💡 Context-Aware: Uses transaction direction (credit/debit) for improved accuracy
• ⚡ Fast Processing: Categorizes 900+ transactions in seconds
• 🎯 High Accuracy: 95%+ overall accuracy with hybrid deep learning + rules approach

---

🎬 Demo

Upload & Categorize

Upload your bank statement (CSV/Excel), and the AI automatically detects columns and categorizes each transaction.

Analytics Dashboard

View interactive charts showing spending distribution, category-wise breakdowns, and top expenses.

PDF Report

Download a professional PDF report with complete expense analysis, charts, and smart recommendations.

---

🧠 How It Works

Hybrid Architecture

TransactAI uses a unique hybrid approach combining:

1. Deep Learning: DistilBERT transformer model fine-tuned on 2000+ synthetic Indian banking transactions
2. Rule-Based NLP: Context-aware rules handle edge cases (e.g., distinguishing "CASH DEPOSIT" from "CASHBACK")
3. Transaction Direction Analysis: Uses withdrawal/deposit columns as additional signals
4. Feature Extraction: Extracts keywords, patterns, and metadata from transaction descriptions

Categorization Pipeline

Bank Statement (CSV/Excel)
         ↓
    File Parser & Data Cleaning
         ↓
  Column Detection & Feature Extraction
         ↓
DistilBERT Model + Context-Aware Rules
         ↓
   Category Prediction + Confidence
         ↓
  Analytics, Insights & PDF Report

---

🚀 Installation

Prerequisites

• Python 3.8 or higher
• 4GB+ RAM (for model loading)
• GPU optional (runs efficiently on CPU)

Setup Instructions

1. Clone the repository

   https://github.com/namanj2003/TransactAI.git
   cd TransactAI

2. Create a virtual environment

   python -m venv venv
   source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install dependencies

   pip install -r requirements.txt

4. Download the trained model

   Place the trained DistilBERT model files in the expense_model_distilbert/ directory:

   expense_model_distilbert/
   ├── config.json
   ├── pytorch_model.bin
   ├── tokenizer_config.json
   ├── vocab.txt
   └── special_tokens_map.json
   

5. Ensure model_config.json exists

   This file should contain category mappings in your project root:

   {
     "categories": ["Food", "Shopping", "Travel", ...],
     "id_map": {"0": "Food", "1": "Shopping", ...}
   }

---

💻 Usage

Running the Application

streamlit run app.py

The web app will automatically open in your browser at http://localhost:8501

Step-by-Step Guide

1. Upload Bank Statement

   • Click "Browse files" and select your bank statement (CSV or Excel format)
   • Supported formats: .csv, .xlsx, .xls

2. Column Mapping

   • The app automatically detects description, withdrawal, deposit, and date columns
   • Adjust mappings manually if needed

3. Categorize Transactions

   • Click "🚀 Categorize Transactions" to run the AI model
   • Progress bar shows real-time processing status

4. View Results

   • Data Tab: View categorized transactions with confidence scores
   • Charts Tab: Interactive visualizations (pie chart, bar chart, top expenses)
   • Insights Tab: Personalized recommendations and financial summary

5. Export Data

   • Download categorized data as CSV
   • Generate professional PDF report with analytics

6. Upload New File

   • Click "🔄 Upload New File" to reset and start fresh

Example Bank Statement Format

Date	Narration	Withdrawal	Deposit
01-01-2024	SWIGGY UPI-4023049214	350.00
02-01-2024	SALARY CREDIT		50000.00
03-01-2024	CASH DEPOSIT ATM		10000.00
04-01-2024	UPI-AMAZON PAY	1250.00

---

📂 Project Structure

transactai/
├── app.py                          # Main Streamlit application
├── model_utils.py                  # Model loading & hybrid prediction logic
├── file_processors.py              # CSV/Excel parsing & cleaning
├── recommendations.py              # Financial insights generation
├── pdf_generator.py                # PDF report creation
├── requirements.txt                # Python dependencies
├── model_config.json               # Category ID mappings
├── expense_model_distilbert/       # Trained DistilBERT model directory
│   ├── config.json
│   ├── pytorch_model.bin
│   └── tokenizer files
└── README.md                       # This file

---

🏗️ Model Architecture

DistilBERT Fine-Tuning

• Base Model: distilbert-base-uncased (66M parameters, 40% smaller than BERT)
• Task: Multi-class text classification (16 expense/income categories)
• Training Data: 2000+ synthetic transaction descriptions covering all major Indian bank patterns
• Fine-tuning: Cross-entropy loss with AdamW optimizer, 3-5 epochs
• Inference: Text → Tokenization → DistilBERT → Softmax → Category prediction

Hybrid Enhancement

The system combines neural predictions with expert rules:

# Simplified hybrid logic
def predict(transaction_text, credit_amount, debit_amount):
    # Step 1: Get DistilBERT prediction
    category, confidence = distilbert_model.predict(transaction_text)
    
    # Step 2: Apply context-aware rules
    if "cash deposit" in text.lower() and credit_amount > 0:
        category = "Income"  # Override neural prediction
    
    if category == "Cashback" and "cashback" not in text.lower():
        category = "Income"  # Fix common misclassification
    
    # Step 3: Use transaction direction
    if credit_amount > 0 and debit_amount == 0:
        if category not in ["Income", "Cashback"]:
            category = "Income"  # Credit must be income-related
    
    return category, confidence

Why Hybrid Approach?

• Neural Network Strength: Understands context and semantic meaning
• Rule-Based Strength: Handles edge cases and banking-specific logic
• Combined Result: 95%+ accuracy vs 87% from DistilBERT alone

---

📊 Results

Performance Metrics

Metric	Value
Overall Accuracy	95.3%
Avg Confidence	89.2%
Inference Speed	~50ms/txn
Categories	16
Model Size	255 MB
Supports Banks	All major Indian banks

Category-wise Performance

Category	Precision	Notes
Food	97.2%	High accuracy
Shopping	94.8%	Good performance
Income	99.1%	Rule-enhanced
Cashback	98.5%	Rule-enhanced
Travel	96.3%	Uber, Ola, IRCTC detected
Bill Payment	95.7%	EMI distinguished

---

🛠️ Technologies Used

Core Technologies

• Frontend: Streamlit
• Deep Learning: PyTorch, Hugging Face Transformers
• NLP Model: DistilBERT (distilbert-base-uncased)

Data Processing

• pandas: Data manipulation and cleaning
• numpy: Numerical operations
• openpyxl, xlrd: Excel file reading

Visualization

• Plotly: Interactive charts in web app
• Matplotlib: Static charts for PDF reports

Other Tools

• ReportLab: PDF generation
• scikit-learn: ML utilities (optional)
• regex: Pattern matching for feature extraction

---

🤝 Contributing

Contributions are welcome! Here's how you can help:

1. Fork the repository
2. Create a feature branch

   git checkout -b feature/AmazingFeature

3. Commit your changes

   git commit -m 'Add some AmazingFeature'

4. Push to the branch

   git push origin feature/AmazingFeature

5. Open a Pull Request

Areas for Contribution

• Add support for more banks
• Improve categorization rules
• Add new expense categories
• Enhance PDF report design
• Add OCR support for scanned PDFs
• Multi-language support

---

📝 License

This project is licensed under the MIT License - see the LICENSE file for details.