📄 Credit Memo Auto-Generator

An AI-powered tool that reads financial documents and produces clear, structured first-draft summaries for credit analysts.

---

🎯 Problem Statement

Credit Memo Auto-Generator

Reads documents → Produces a clear first draft

What it is

A tool that reads long financial documents and writes a clear first-draft summary for humans.

In simple terms

AI reads PDFs, pulls key numbers and risks, and produces a structured, editable draft.

Think of it like

ChatGPT that reads PDFs and writes drafts.

---

� Quick Setup

Prerequisites

• Python 3.9 or higher
• OpenAI API key

Installation Steps

1. Clone the repository

   git clone <repository-url>
   cd Memo_Test

2. Install dependencies

   cd backend
   pip install -r requirements.txt

3. ⚠️ IMPORTANT: Set up environment variables

   Create a .env file inside the backend folder (not in the root directory):

   # Navigate to backend folder
   cd backend
   
   # Create .env file
   # Windows (PowerShell)
   New-Item .env
   
   # macOS/Linux
   touch .env

   Add your OpenAI API key to the .env file:

   OPENAI_API_KEY=your_openai_api_key_here

   Note: The .env file must be in Memo_Test/backend/.env for the application to work correctly.

4. Run the application

   # Make sure you're in the backend folder
   cd backend
   streamlit run frontend.py

5. Access the app

   Open your browser and navigate to the port where it is running ( will be shown in the shell)

---

🧪 Test Files

Sample test files are available in the backend/uploads/ folder:

File	Description
final_testcase.pdf	Main test document - standard financial PDF
test3.pdf	Password-protected document (for testing password feature)

Note: When testing test3.pdf, check the "🔒 PDF is password protected" checkbox and enter the password.

---

💡 The Challenge

Credit analysts and financial professionals spend significant time manually reviewing financial documents (annual reports, balance sheets, income statements) to:

• Extract key metrics and trends
• Identify potential risks
• Generate executive summaries for stakeholders

This process is:

• Time-consuming: Manual review of lengthy documents
• Error-prone: Human oversight can miss critical details
• Inconsistent: Different analysts may interpret data differently
• Expensive: Requires skilled professionals for thorough analysis

Our Solution

This tool automates the financial document analysis workflow:

Input	Process	Output
Upload a PDF	AI extracts & analyzes	Structured, editable draft

Example builds:

• ✅ Upload a PDF and generate a one-page executive summary
• ✅ Highlight key numbers and show the source page
• ✅ Export the draft to Markdown or Word for editing

---

📋 Example Output

Upload a PDF with multiple financial statements → tool outputs:

Component	Description
5-bullet executive summary	Key highlights from the document
Key metrics table	Financial metrics with trends (🟢↑ / 🔴↓)
"Top 3 risks" section	Identified risks with severity ratings
Source tracing	Highlight a sentence → see which PDF page it came from
Confidence tags	✅ Strong data | ⚠️ Incomplete data

---

🔄 User Flow

Upload PDF → Click "Generate Memo" → Review → Edit → Download

1. Upload PDF - Drag & drop or select financial document
2. Generate Memo - AI extracts data and generates summary
3. Review - Check executive summary, risks, and metrics
4. Validate - Secondary AI validates accuracy with confidence scores
5. Edit - Refine using chat Q&A or regenerate with feedback
6. Download - Export to Markdown or Text for further editing

---

✨ Features

📊 PDF Data Extraction

• Table Extraction: Automatically detects and extracts financial tables with headers and data
• Text Extraction: Extracts paragraphs with section headings and page references
• Password Support: Handles password-protected PDF documents

🤖 AI-Powered Analysis

Key Metrics Identification

• Automatically identifies key financial metrics from extracted tables
• Shows trends (🟢 Increase / 🔴 Decrease) for each metric
• Provides explanations with source references

Executive Summary Generation

• Generates configurable number of summary bullet points (3-10)
• Categories: Financial Performance, Revenue, Profitability, Cash Flow, etc.
• Confidence indicators (High/Medium/Low) for each point
• Page references for traceability

Key Risks Assessment

• Identifies and categorizes financial risks
• Severity ratings (High/Medium/Low) with strict criteria
• Evidence-based risk descriptions
• Categories: Liquidity, Credit, Operational, Compliance, etc.

🔒 Privacy & Security

Pseudonymization

• Optional data anonymization before sending to LLM
• Replaces company names, people, and products with tokens
• Automatically restores original names in final output
• Configurable entity mappings

💬 Document Q&A Chat

• Conversational interface for querying document data
• Context-aware responses based on extracted content
• Suggested questions for quick exploration
• Conversation history for multi-turn interactions

✅ Secondary Validation

• Uses a different AI model (GPT-4o) to validate primary outputs
• Scores each summary point and risk on:
  • Factual Grounding (1-5)
  • Numeric Accuracy (1-5)
  • Coherence (1-5)
• Identifies critical issues: Hallucinations, Numeric Errors, Unsupported Claims
• Provides correction suggestions

🔄 Feedback-Driven Regeneration

• Regenerate memos incorporating validation feedback
• LLM uses identified issues to produce improved output
• Iterative refinement workflow

💰 Token Usage Monitoring

• Real-time tracking of API token consumption
• Cost estimation based on model pricing
• Breakdown by model and service
• Session history and downloadable logs

📥 Export Reports

• Generate comprehensive Markdown reports
• Includes executive summary, risks, and key metrics
• Download as .md or .txt files
• Professional formatting with trend indicators

🖥️ Local LLM Support

• Alternative branch with Ollama integration
• Runs inference locally without cloud API calls
• Same functionality as cloud-based analysis

---

🛠️ Tech Stack

Component	Technology
Frontend	Streamlit
PDF Processing	pdfplumber
Data Handling	pandas, numpy
LLM Provider	OpenAI (GPT-4o-mini, GPT-4o)
Environment	python-dotenv

---

Project Structure

Memo_Test/
├── README.md
└── backend/
    ├── frontend.py              # Streamlit UI application
    ├── requirements.txt         # Python dependencies
    ├── .env                     # Environment variables (create this)
    ├── app/
    │   ├── __init__.py
    │   └── services/
    │       ├── __init__.py
    │       ├── pdf_table_extractor.py   # Table extraction service
    │       ├── pdf_text_extractor.py    # Text extraction service
    │       ├── pseudonymizer.py         # Data anonymization service
    │       ├── llm_insights.py          # Key metrics analysis
    │       ├── generate_memo.py         # Memo generation service
    │       ├── chat_service.py          # Document Q&A chat
    │       ├── secondary_validator.py   # Output validation service
    │       └── token_monitor.py         # Token usage tracking
    ├── config/
    │   ├── entity_map.json      # Pseudonymization mappings
    │   └── mapping_audit.json   # Audit trail for mappings
    ├── uploads/                 # Uploaded PDF storage
    └── logs/                    # Token usage logs

---

🚀 Usage Guide

Basic Workflow

1. Upload a PDF: Select a financial document (supports password-protected files)

2. Generate Summary: Click "Generate Summary" to extract tables and text

3. View Extracted Data:

   • Tables Tab: View extracted financial tables
   • Text Tab: View extracted text paragraphs

4. Generate Memo:

   • Go to "AI Insights" tab
   • Configure number of summary points and risks
   • Optionally enable pseudonymization
   • Click "Generate Memo"

5. Validate Output:

   • Go to "Validation" tab
   • Click "Validate Memo" to check accuracy
   • Review scores and critical issues

6. Iterate if Needed:

   • Use "Regenerate Memo" to improve based on feedback

7. Export Report:

   • Go to "Export" tab
   • Download as Markdown or Text

Advanced Features

• Chat Q&A: Use the Chat tab to ask specific questions about the document
• Key Metrics: View extracted metrics with trend analysis in the Key Metrics tab
• Token Monitoring: Track API usage and costs in the Token Usage tab

---

⚙️ Configuration

Entity Mappings (Pseudonymization)

Edit backend/config/entity_map.json to customize entity replacements:

{
  "companies": {
    "Acme Corp": "COMPANY_001"
  },
  "people": {
    "John Smith": "PERSON_001"
  }
}

Model Pricing (Token Monitor)

Edit backend/app/services/token_monitor.py to update pricing:

MODEL_PRICING = {
    "gpt-4o": {"input": 0.005, "output": 0.015},
    "gpt-4o-mini": {"input": 0.00015, "output": 0.0006},
    # Add more models...
}

---

🔮 Future Enhancements

• Batch processing for multiple documents
• Custom LLM model selection
• Integration with document management systems
• Advanced visualization dashboards
• API endpoints for programmatic access
• Multi-language support

---

👥 Team

• Document analysis and extraction pipeline
• LLM integration and prompt engineering
• Frontend development and UX
• Local LLM support (Ollama integration - separate branch)

---

📄 License

This project is for educational and demonstration purposes.

---

🙏 Acknowledgments

• OpenAI for GPT models
• Streamlit for the rapid UI framework
• pdfplumber for PDF processing capabilities