TIARA - Tata Intelligent Agent for Real-world Agents

An intelligent multi-agent AI system revolutionizing personal loan sales through conversational banking

Built for EY Techathon 6.0 - BFSI Challenge

🎥 Demo Video

👇 Click the thumbnail image to watch video on youtube

📋 Table of Contents

Overview
Problem Statement
Solution Architecture
Agentic AI Workflow
Key Features
Tech Stack
Installation & Setup
User Journey
Underwriting Logic
Test Scenarios
Project Structure
Team
Contact

Overview

TIARA (Tata Intelligent Agent for Real-world Agents) is a next-generation conversational AI system that transforms the personal loan sales process. Built on a multi-agent architecture, TIARA orchestrates specialized AI agents to deliver a seamless, human-like loan application experience from initial conversation to sanction letter generation.

Challenge Context

EY Techathon 6.0 - BFSI Track: Tata Capital

Tata Capital, a leading NBFC, aims to increase personal loan sales through an AI-driven web chatbot. The challenge requires building an agentic AI solution where a Master Agent coordinates multiple Worker Agents to handle the complete loan journey - from discovery to approval.

Problem Statement

Business Objective

Improve personal loan sales success rate through an AI-driven conversational approach that:

Engages customers landing via digital ads/marketing emails
Understands customer needs and convinces them to take personal loans
Completes end-to-end process: verification → underwriting → sanction letter generation

Key Requirements

Master Agent: Orchestrates conversation flow and coordinates worker agents
Worker Agents:
- Sales Agent: Negotiates terms, discusses amount/tenure/rates
- Verification Agent: Confirms KYC details and document uploads
- Underwriting Agent: Validates eligibility based on credit rules
- Sanction Generator: Creates automated PDF approval letters

Solution Architecture

              ┌─────────────────────────────────────────────────────────────────┐
              │                          TIARA System                           │
              │                   (Master Agent Controller)                     │
              │              Powered by Google Gemini 2.5 Flash                 │
              └────────────────────────┬────────────────────────────────────────┘
                                       │
                      ┌────────────────┼────────────────┐
                      │                │                │
                      ▼                ▼                ▼
              ┌───────────────┐ ┌──────────────┐ ┌──────────────┐
              │ Discovery     │ │ Sales Agent  │ │ Verification │
              │ & Welcome     │ │ (Negotiation)│ │ Agent (KYC)  │
              │ (Master)      │ │              │ │              │
              └───────────────┘ └──────────────┘ └──────────────┘
                      │                │                │
                      └────────────────┼────────────────┘
                                       ▼
                      ┌────────────────────────────────┐
                      │    Underwriting Agent          │
                      │    (Credit Evaluation)         │
                      └────────────────┬───────────────┘
                                       ▼
                      ┌────────────────────────────────┐
                      │   Sanction Letter Generator    │
                      │   (PDF Generation - pdf-lib)   │
                      └────────────────────────────────┘

Agentic AI Workflow

Stage-by-Stage Journey

Stage 1: Welcome & Authentication

Master Agent greets the customer
Phone number collection & OTP verification
Customer profile retrieval from mock CRM

Stage 2: Discovery

Master Agent understands loan requirements
Extracts: Purpose, Amount, Preferred Tenure
Natural language understanding for amount detection (₹3 lakhs, 300000, etc.)

Stage 3: Sales

Sales Agent presents personalized loan offer
Shows EMI calculations with tenure options (36/48/60 months)
Handles negotiation and customer objections
Displays credit score and affordability ratios

Stage 4: Verification

Verification Agent requests KYC documents
Document upload interface: Aadhaar, PAN, Salary Slip (conditional)
Real-time upload validation

Stage 5: Underwriting

Underwriting Agent evaluates eligibility
Credit score verification (min 700/900)
Pre-approved limit checks
EMI-to-salary ratio validation (max 50%)

Stage 6: Sanction

Sanction Generator creates PDF offer letter
Instant download with loan details
Unique Loan ID generation

Key Features

🎭 Multi-Agent Orchestration

Master Agent seamlessly coordinates 4 specialized worker agents
Intelligent stage transitions based on conversation context
Agent handoff notifications for transparency

💬 Human-Like Conversation

Natural language understanding using Google Gemini 2.5 Flash
Context-aware responses with conversation memory
Smart amount/tenure extraction from free-form text

📊 Intelligent Underwriting

Approval Rules:
✓ Amount ≤ Pre-approved Limit → Instant Approval
✓ Amount ≤ 2× Limit + Salary Slip → Conditional Approval
✗ Amount > 2× Limit → Rejection
✗ Credit Score < 700 → Rejection
✗ EMI > 50% Salary → Rejection

🌐 Multi-Language Support

English, Hindi, Odia, Tamil, Malayalam, Marathi, Kannada, Bengali
Language switcher in chat interface

📄 Automated Sanction Letters

PDF generation with loan details
Downloadable offer letters
Professional formatting with Tata Capital branding

Tech Stack

Frontend

Next.js 16 - React framework with App Router
TypeScript - Type-safe development
Tailwind CSS - Utility-first styling
Lucide Icons - Modern icon library

Backend

Next.js API Routes - Serverless functions
OpenRouter API - AI model orchestration
Google Gemini 2.5 Flash Lite - Conversational AI

Document Generation

PDFKit - PDF creation library
Canvas-based rendering for professional documents

State Management

React Hooks (useState, useEffect, useRef)
localStorage - Session persistence
Server-side memory store for conversation history

Installation & Setup

Prerequisites

Node.js 18.x or higher
npm or yarn package manager

Clone Repository

git clone https://github.com/sujit-prog/EY.git
cd EY

Install Dependencies

npm install
# or
yarn install

Environment Variables

Create a .env.local file:

OPENROUTER_API_KEY=your_openrouter_api_key_here

Run Development Server

npm run dev
# or
yarn dev

Open http://localhost:3000 in your browser.

Build for Production

npm run build
npm start

User Journey

Step-by-Step Flow

Landing
- Customer arrives via marketing campaign
- TIARA welcomes with friendly greeting
Authentication
- Phone number input: 9337236782
- OTP verification (any 4-digit code accepted in demo)

Conversation

User: "I need a loan for education"
TIARA: "Great! How much are you looking to borrow?"

User: "3 lakhs"
TIARA: "Based on your profile, we can offer ₹3 lakhs at 10.5% p.a..."

Negotiation
- EMI calculations shown
- Tenure options presented
- Customer can adjust terms
Document Upload
- Aadhaar Card ✓
- PAN Card ✓
- Salary Slip (if amount > pre-approved limit) ✓
Instant Decision
- Underwriting completes in seconds
- Approval/rejection notification
Sanction Letter
- PDF download with full loan details
- Unique Loan ID for reference

Underwriting Logic

Approval Matrix

Scenario	Amount	Credit Score	Salary Slip	EMI Ratio	Decision
Instant	≤ Pre-approved	≥ 700	Not Required	≤ 50%	✅ APPROVED
Conditional	≤ 2× Limit	≥ 700	Required	≤ 50%	✅ APPROVED
High Amount	> 2× Limit	Any	Any	Any	❌ REJECTED
Low Credit	Any	< 700	Any	Any	❌ REJECTED
High EMI	Any	≥ 700	Yes	> 50%	❌ REJECTED

Example Calculations

Customer: Dharmendra Mahanta
Salary: ₹60,000/month
Pre-approved: ₹4.5 lakhs
Credit Score: 780/900

Scenario 1: ₹3 lakhs request
→ Within limit → Instant Approval ✅

Scenario 2: ₹7 lakhs request  
→ < 2× limit (₹9L) → Needs salary slip
→ EMI: ₹16,500 (27.5% of salary) → Approved ✅

Scenario 3: ₹10 lakhs request
→ > 2× limit → Rejected ❌

Test Scenarios

Pre-Configured Test Users

Phone	Name	Salary	Credit	Pre-approved	Scenario
`9876543210`	Rahul Sharma	₹75K	820	₹6L	Instant Approval
`9337236782`	Dharmendra Mahanta	₹60K	780	₹4.5L	Conditional
`9876543214`	Vikram Singh	₹55K	650	₹2L	Low Credit Score
`9876543212`	Amit Patel	₹125K	850	₹10L	Premium Customer

Testing Flows

Flow 1: Instant Approval

Phone: 9876543210
Amount: ₹4 lakhs
Expected: Approved without salary slip

Flow 2: Salary Slip Required

Phone: 9337236782  
Amount: ₹7 lakhs
Expected: Approved after salary slip upload

Flow 3: Rejection - Amount

Phone: 9876543216
Amount: ₹10 lakhs
Expected: Rejected (exceeds 2× limit)

Project Structure

EY/
├── app/
│   ├── page.tsx                 # Landing page
│   ├── layout.tsx               # Root layout
│   ├── api/
│   │   └── chat/
│   │       └── route.ts         # Multi-agent orchestration logic
│   └── loan/
│       └── chat/
│           └── page.tsx         # Chatbot page route
├── components/
│   └── chat/
│       ├── ChatShell.tsx        # Main chat UI component
│       ├── MessageBubble.tsx    # Individual message rendering
│       ├── SuggestionChips.tsx  # Message suggestion
│       └── StageIndicator.tsx   # Visual stage progress
├── lib/
│   ├── users.ts                 # Mock CRM data (10 customers)
│   ├── prompts.ts               # Agent system prompts
│   ├── emi.ts                   # EMI calculation utilities
│   ├── pdf.ts                   # Sanction letter generator
│   ├── memory.ts                # Session state management
│   ├── loanConfig.ts            # Interest rates & terms
│   └── openrouter.ts            # AI model configuration
├── public/                      # Static assets
├── .env.local                   # Environment variables
└── README.md                    # This file

Team

Team Name: Veggies

Team Members:

Name	Role	GitHub
Dharmendra Mahanta	Backend Development	@dharmendra-007
Sujit Kumar Sha	Frontend Development	@sujit-prog
Alpana Mohanty	Agentic AI Backend Development	@Al-Pa-Na
Aditi Panda	Backend,system integration	@aditipanda01
Prerana Priyadarsini Das	Business Logic & Rule Design	@preranadas03

Institution: Veer Surendra Sai University Of Technology, Burla, Sambalpur

Hackathon: EY Techathon 6.0 - BFSI Challenge

Key Learnings

Agentic AI Design Principles

Clear Agent Boundaries: Each agent has a specific, well-defined role
Master Orchestration: Central controller manages workflow transitions
Context Preservation: Conversation history maintained across agents
Error Handling: Graceful degradation when agents fail
Human-in-the-Loop: User controls pace of conversation

Technical Achievements

✅ Implemented multi-agent orchestration with 4 specialized AIs
✅ Built intelligent underwriting with 5+ business rules
✅ Created natural language understanding for loan parameters
✅ Developed real-time document upload simulation
✅ Generated professional PDF sanction letters
✅ Achieved sub-2-second response times

Edge Cases Handled

Edge Case	Handling Strategy
User uploads wrong document	Re-upload prompt with clear instructions
Amount in different formats	Regex patterns for lakhs/L/numbers
Tenure not specified	Default to 48 months, allow adjustment
Network failure	Error message with retry option
Invalid phone number	Validation before OTP generation
Borderline credit score	Clear rejection message with reason
Multiple loan requests	Session isolation via unique IDs

Future Enhancements

Voice Integration: Speech-to-text for hands-free interaction
Live Credit Bureau API: Real-time credit score fetching
Co-applicant Support: Joint loan applications
EMI Calculator Widget: Interactive affordability tool
WhatsApp Integration: Loan application via messaging
Predictive Analytics: ML-based approval likelihood
Regional Language NLP: Advanced multilingual support
Video KYC: Live verification for high-value loans

Contact

For queries regarding this project:

Email: dev.dharmendra.m@gmail.com
LinkedIn: @dharmendram007
GitHub: @dharmendra-007

Built with ❤️ for EY Techathon 6.0

Empowering India's Digital Future Through Agentic AI

Name		Name	Last commit message	Last commit date
Latest commit History 27 Commits
app		app
components/chat		components/chat
lib		lib
public		public
ui		ui
.gitignore		.gitignore
README.md		README.md
eslint.config.mjs		eslint.config.mjs
next.config.ts		next.config.ts
package-lock.json		package-lock.json
package.json		package.json
postcss.config.mjs		postcss.config.mjs
tsconfig.json		tsconfig.json

Folders and files

Latest commit

History

Repository files navigation