Agentic Loan Assistant Chatbot
EY Hackathon 2025 — Agentic loan assistant chatbot.
Overview
An AI-powered loan origination system that lets users apply for a loan through either a website chat or WhatsApp. Both channels feed into a single Master Agent, which orchestrates verification, risk assessment, underwriting and document generation, and returns the outcome (including the sanction letter) on whichever channel the user started from.
Demo video: YouTube
What it does
- Website chat and WhatsApp chat support
- AI-driven conversation using a Groq LLM
- KYC and document verification
- Credit and risk assessment using XGBoost
- Automated loan approval or rejection
- Sanction letter PDF generation, delivered via WhatsApp or as a website download
- Multilingual support via Deeptranslator
The WhatsApp user flow, as described in the README:
- User sends "Hi"; the bot creates a new session
- Bot asks for a phone number and validates it
- Bot collects loan amount, loan purpose and monthly income
- Bot requests KYC documents, which are then verified
- Credit score is fetched and the XGBoost risk model is executed
- Loan is approved or rejected; if approved, a sanction letter PDF is generated and sent via WhatsApp, with a website download link provided as well
How it's built
High-level architecture:
- Website Chat → Flask API
- WhatsApp → Twilio → Flask webhook
- Flask → Master Agent → specialized agents → backend systems
- Output → website / WhatsApp
Stack:
| Layer | Technology |
|---|---|
| Backend | Flask (Python) |
| AI | Groq LLM |
| Risk model | XGBoost |
| Messaging | Twilio WhatsApp API |
| Tunneling | Ngrok |
| PDF generation | ReportLab |
| Document parsing | OCR / NLP |
| Frontend | React + TypeScript (Vite) |
| Authentication | Firebase |
| Multilingual | Deeptranslator |
For WhatsApp, a separate Flask server (whatsapp_bot.py) runs on port 5000, is exposed publicly through Ngrok, and its /webhook URL is registered in the Twilio WhatsApp Sandbox.
Context
The README documents the architecture, the stack and the WhatsApp flow above. The project was built for the EY Hackathon 2025; the README itself records no competition, so that attribution comes from the owner rather than from the repository.