FinGuard: The Mis-selling Radar

A Proactive AI Tool for Supervisory Technologies (SupTech)

Overview

FinGuard is a decentralized "SupTech" (Supervisory Technology) prototype designed to assist financial regulators in monitoring mis-selling risks in real-time.

Financial products often suffer from a gap between "Expectation" (marketing promises) and "Reality" (customer experience). FinGuard bridges this gap by autonomously comparing product documentation against unstructured consumer feedback.

The Goal: Move regulatory oversight from reactive (waiting for formal complaints) to proactive (flagging discrepancies as they emerge).

Core Architecture

FinGuard operates on a Dual-Pipeline Architecture that converges into a central Gap Analyzer.

1. The Expectation Engine (Pipeline A):

• Input: Product PDFs (Brochures, Scheme Information Documents).
• Tech: OCR, NLP (LLMs/LangChain).
• Function: Extracts semantic "Promises" (e.g., "Guaranteed Returns," "Zero Exit Load").
• Output: A structured JSON Promise_Profile.

2. The Reality Engine (Pipeline B):

• Input: Unstructured text (Social media, reviews, forums).
• Tech: Sentiment Analysis, Topic Modeling (BERTopic/RoBERTa).
• Function: Extracts "Lived Experiences" and maps sentiment to specific product features.
• Output: A Customer_Dissatisfaction_Index.

3. The Gap Analyzer:

• Function: Calculates vector distances and semantic discrepancies between Promises and Reality.
• Result: A Mis-selling Risk Score (0-100).

Key Features

• Automated Promise Extraction: Parses complex financial PDFs to identify key claims regarding Lock-in periods, Risk profiles, and Fees without manual entry.
• Aspect-Based Sentiment Analysis: Doesn't just tell you if a review is "sad"; it tells you why (e.g., "Positive on Returns" but "Negative on Hidden Fees").
• The Fine-Print Detector: Compares prominent marketing text against obscure footnotes to flag "visual obfuscation."
• Real-time Dashboard: A Streamlit-based UI for regulators to view Risk Scores and drill down into specific product discrepancies.

Technology Stack

| Component | Technology | Role | | | | | | Language | Python 3.10+ | Core Application Logic | | Orchestration | LangChain | LLM Framework for Promise Extraction | | LLM Provider | OpenAI API / Local Llama | Zero-shot Entity Extraction | | NLP & Topic Modeling | BERTopic, Spacy | Clustering Customer Feedback | | Vector Store | ChromaDB / FAISS | Semantic Similarity Search | | Frontend | Streamlit | Interactive Regulatory Dashboard | | Data Processing | Pandas, PyMuPDF | Data Manipulation & PDF Parsing |

Quick Start

Prerequisites

• Python 3.10 or higher
• An OpenAI API Key (or a local LLM endpoint setup)

Installation

1. Clone the Repository

git clone https://github.com/yourusername/finguard.git
cd finguard

2. Create Virtual Environment

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

3. Install Dependencies

pip install -r requirements.txt

4. Configuration Create a .env file in the root directory:

OPENAI_API_KEY=sk-your-key-here
DATA_SOURCE_PATH=./data

Running the Application

To launch the regulatory dashboard:

streamlit run app.py

Project Structure

finguard/
├── assets/              # Simulated PDF brochures & review datasets
├── src/
│   ├── engines/
│   │   ├── expectation.py  # PDF Parsing & Extraction Logic
│   │   └── reality.py      # Sentiment & Topic Modeling Logic
│   ├── analyzer/
│   │   └── gap_logic.py    # Comparison Algorithms
│   └── utils/
│       └── data_loader.py
├── app.py               # Main Streamlit Dashboard entry point
├── requirements.txt     # Python dependencies
├── README.md            # Project documentation
└── COOKBOOK.md          # Developer Guide & User Flows

Team

• [Nishit Somani] - AI Engineering and Optimization (Kaptaan Sahab)
• [Vedant Hingu] - Frontend Design and Database organization
• [Hetansh Panchal] - Database Enginnering and Optimization
• [Neel Patel] - Backend monitoring and tool control
• [Utsav Mistry] - Backend impplememntation and Architecture

Built with ❤️ and ☕ at [DA-IICT].