GlassScore

Solution Overview

Problem: Traditional credit scoring systems are often "black boxes," relying solely on rigid financial history and failing to capture the full picture of an applicant's trustworthiness. They struggle with thin-file applicants and lack transparency.

Target User: Financial institutions, loan officers, and credit analysts seeking a more holistic and explainable risk assessment tool.

Fix: GlassScore is an advanced, multi-modal credit scoring platform that combines traditional Machine Learning with Large Language Models (LLMs) and real-time Web Search. It provides a transparent, explainable, and comprehensive credit score by analyzing structured data, personal statements, and external digital footprints.

Demo Video

Watch on YouTube

Slides

GlassScore Presentation (PDF) | View on Canva

Snapshots

Landing page

Info page

Fill applicant data Upload related docs

Evaluation

Streaming Evidence

Clear citation

LLM Summary

Evidence Override Protocol

Getting Started

Prerequisites

• Python 3.10+
• Node.js 18+
• API Keys for OpenAI/Google Gemini and Tavily.

Installation

1. Clone the repository:

   git clone https://github.com/stanx19/GlassScore.git

2. Navigate to the root directory:

   cd GlassScore

Backend Setup

1. Navigate to the backend directory:

   cd backend

2. Create and activate a virtual environment:

   python -m venv venv

   Windows

   .\venv\Scripts\activate

   Linux/Mac

   source venv/bin/activate

3. Install dependencies:

   pip install -r requirements.txt

4. Set up environment variables:

   cp .env.example .env

   Remember to fill in the API keys for OpenAI/Google Gemini and Tavily accordingly.

5. Run the server:

   uvicorn main:app --reload

Frontend Setup

1. Navigate to the frontend directory:

   cd frontend/GlassScore

2. Install dependencies:

   npm install

3. Start the development server:

   npm run dev

Tech Stack

• Frontend: React, TypeScript, Vite, Axios, Tailwind CSS.
• Backend: Python, FastAPI, Uvicorn.
• AI & ML: LangChain, OpenAI GPT-4 / Google Gemini, Scikit-learn, SHAP, Tavily Search API.
• Database: PostgreSQL (Production) / In-Memory (Dev).

System Architecture

GlassScore operates on a modern client-server architecture:

Data Flow:

1. Input: User submits loan application data (structured) and supporting documents/statements (unstructured).
2. Processing:
   • ML Engine: Predicts default risk using historical loan data.
   • LLM Engine: Analyzes text for behavioral risk markers (e.g., gambling references, inconsistencies).
   • Web Engine: Verifies claims against live web data (e.g., employment verification).
3. Output: Real-time stream of "Evidence" points (Positive/Negative) that dynamically adjust the GlassScore.

Key Features

• Multi-Modal Evaluation: Seamlessly integrates structured financial data (Random Forest Model), unstructured text analysis (LLM), and real-time web verification.
• Pipeline: Automated preprocessing pipeline including median imputation for numerical values and one-hot encoding for categorical variables.
• Real-Time Streaming: Results are pushed to the frontend instantly as they are processed, providing immediate feedback.
• Explainable AI (XAI): Every score adjustment is backed by specific evidence, citations, and reasoning.
• Interactive Dashboard: A modern, responsive UI that visualizes the scoring process and allows deep dives into specific risk factors.
• Dynamic Risk Assessment: Adapts to new information found during the web search and text analysis phases.

Machine Learning Model

GlassScore employs a robust Random Forest Classifier to predict loan default risk based on historical data.

• Algorithm: Random Forest Classifier (Scikit-learn).
• Training: Optimized using RandomizedSearchCV with 3-fold cross-validation for hyperparameter tuning.
• Features:
  • Numerical: Age, Income, Employment Length, Loan Amount, Interest Rate, Loan/Income Ratio, Credit History Length.
  • Categorical: Home Ownership, Loan Intent, Loan Grade, Default History.
• Explainability: Integrated SHAP (SHapley Additive exPlanations) values to provide local interpretability, explaining exactly why a specific applicant received a certain risk score (e.g., "High Loan Amount" or "Low Income").

Innovation & Differentiation

• Beyond the FICO Score: Moves beyond simple number crunching to understand the context of a borrower.
• Agentic Verification: Uses autonomous agents to verify user claims on the web, reducing fraud risk.
• Transparent Decisioning: Unlike traditional models, GlassScore tells you why a decision was made.