🚀 Internship Project | Tiered AI Architecture | BFSI Domain

BFSI Call Center AI Assistant

Overview

This project is a Tiered AI Assistant for BFSI (Banking, Financial Services, Insurance) queries built using:

• FastAPI

• Sentence Transformers

• FAISS

• DistilGPT2 (Fine-tuned SLM)

• Retrieval-Augmented Generation (RAG)

The system is designed to provide safe, compliant, and structured financial responses using a multi-tier architecture.

Architecture

Tiered Flow

User Query

  │

  ▼

Guardrails (Sensitive Filtering)

  │

  ▼

Dataset Similarity (FAISS Search)

  │

  ├── Match Found → Return Dataset Response

  │

  ▼

Complex Query?

  │

  ├── YES → RAG (Policy Retrieval + SLM)

  │

  └── NO → SLM Fallback

Key Features

• Dataset-first compliance

• Sensitive information guardrails

• Semantic similarity search

• Retrieval-Augmented Generation (RAG)

• Fine-tuned Small Language Model (SLM)

• Tiered fallback architecture

Project Structure

bfsi_ai_assistant/

│

├── main.py

├── services/

│ ├── guardrails.py

│ ├── similarity.py

│ ├── rag.py

│ └── model.py

│

├── data/

│ ├── alpaca_dataset.json

│ └── knowledge_docs/

│ └── policies.txt

│

├── train_slm.py

├── generate_dataset.py

├── requirements.txt

└── README.md

Setup Instructions

1. Clone Repository

git clone https://github.com/your-username/bfsi\_ai\_assistant.git

cd bfsi_ai_assistant

2. Create Virtual Environment

python -m venv venv

venv\Scripts\activate # Windows

3. Install Dependencies

pip install -r requirements.txt

4. Train SLM (Optional)

python train_slm.py

5. Run Application

uvicorn main:app --reload

Open:

http://127.0.0.1:8000/docs

Guardrails

The system blocks:

• Account numbers

• OTPs

• Passwords

• Credit card details

• Crypto queries

• Sensitive matters and information

Future Improvements

• LoRA fine-tuning

• SME-reviewed dataset expansion

• Policy auto-update pipeline

• Multilingual support