AI-Powered Loan Approval System: From Explainability to Intelligence

This repository demonstrates a comprehensive approach to building an intelligent loan approval system, progressing from model interpretability to an interactive RAG-based assistant.

Part 1: Model Interpretability

The first part focuses on understanding and explaining AI model predictions in the loan approval context using state-of-the-art interpretability techniques.

Features

• LIME (Local Interpretable Model-agnostic Explanations) for individual predictions
• SHAP (SHapley Additive exPlanations) for feature importance analysis
• Comparison with inherently interpretable models like decision trees
• Comprehensive visualizations

Setup & Usage

# Create and activate virtual environment
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

# Start Jupyter Notebook
jupyter notebook

Navigate to model_interpretability/model_explanations.ipynb to explore the demonstrations.

Part 2: Intelligent Loan Approval Assistant

Building upon our understanding of model interpretability, we've developed an interactive loan approval assistant using Retrieval Augmented Generation (RAG) technology.

Features

• Interactive Streamlit web interface
• RAG-based decision explanation using domain-specific knowledge
• Real-time loan approval decisions
• Detailed criteria analysis and explanations
• Integration with Microsoft's Phi-2 language model

Demo

▶️ Download the original demo video (.webm)

Setup & Usage

# Ensure you're in the virtual environment
cd loan_approval_rag
pip install -r requirements.txt

# Run the Streamlit app
streamlit run loan_approval_rag/src/app.py --server.fileWatcherType none

Architecture

The architecture of the RAG-based Intelligent Loan Approval Assistant is as follows:

+----------------------+     +------------------+     +----------------------+
|    User Interface    |     |   RAG Pipeline   |     |   Domain Knowledge   |
|      (Streamlit)     |<--->|   Coordinator    |     |   Loan Documents     |
+----------------------+     +------------------+     +----------------------+
                                     |                           |
                                     v                           v
                            +------------------+     +----------------------+
                            |    Retriever     |<----| Document Indexing &  |
                            | FAISS + Sentence |     | Embedding Process   |
                            |   Transformers   |     +----------------------+
                            +------------------+
                                     |
                                     v
                            +------------------+
                            |    Generator     |
                            |    (Phi-2 LLM)   |
                            +------------------+
                                     |
                                     v
                            +------------------+
                            |  Final Answer    |
                            | Loan Decision &  |
                            |   Explanation    |
                            +------------------+

Architecture (Mermaid Diagram)

flowchart TD
    style UI fill:#e3f6fc,stroke:#38b6ff,stroke-width:2px
    style FILES fill:#f9f9f9,stroke:#bdbdbd,stroke-width:2px
    style LLM fill:#fffbe7,stroke:#f9d923,stroke-width:3px
    style DOCS fill:#f0e6fa,stroke:#a259d9,stroke-width:2px

    UI([👤<br/>User<br/>Interface])
    FILES([🗂️<br/>User<br/>Files])
    DOCS([📄<br/>Domain<br/>Knowledge])
    LLM([🧠<br/>AI Model])

    %% Step flows
    UI -- "1️⃣ Provide Files" --> FILES
    FILES -- "2️⃣ Indexed & Embedded" --> DOCS
    UI -- "3️⃣ Application & Questions" --> LLM
    DOCS -- "4️⃣ Relevant Info" --> LLM
    LLM -- "5️⃣ Decision & Explanation" --> UI

Loading

The diagram shows the key components and data flow of our RAG-based loan approval system:

1. Domain Knowledge: Loan documents are indexed and embedded
2. Retrieval: FAISS retrieves relevant context based on user queries
3. Generation: Phi-2 LLM generates responses using retrieved context
4. Pipeline: Coordinates the flow between components
5. User Interface: Streamlit web app for user interaction

Project Evolution

This repository showcases the evolution of AI in loan approval systems:

1. Understanding the Black Box: We start with model interpretability, making AI decisions transparent and trustworthy through LIME and SHAP analysis.

2. Enhancing with Intelligence: We then progress to a RAG-based system that not only makes decisions but also provides context-aware explanations by leveraging domain knowledge and large language models.

Technical Stack

Model Interpretability

• LIME & SHAP for model explanations
• Scikit-learn for traditional ML models
• Matplotlib & Seaborn for visualizations

RAG-based Assistant

• LangChain for RAG pipeline
• Microsoft Phi-2 for language modeling
• Streamlit for web interface
• FAISS for vector storage
• Sentence Transformers for embeddings

Citation

If you use or reference this project as part of academic work or publications, please cite the following chapter from the book project:

@incollection{Lu2026Computer,
  author    = {Tianxiang Lu},
  title     = {Den Computer verstehen: Von Erklärbarkeit zu automatisiertem Reasoning},
  booktitle = {Technologie und Organisation},
  editor    = {Cornelia Reindl and Bernhard Wecke and Ulrich Irnich},
  publisher = {Springer Gabler, Wiesbaden},
  year      = {2026},
  pages     = {17--32},
  doi       = {10.1007/978-3-658-51511-9_2},
  url       = {https://doi.org/10.1007/978-3-658-51511-9_2}
}

Note

This repository is part of a book project demonstrating the practical evolution of AI systems, from explainable models to intelligent assistants. Each component is designed to showcase best practices in both technical implementation and user experience.