• 

  CCCD scanning function, scan to upload personal information

• 

  AI Loan Simulation for Customer, Dealer Sales

• 

  Scan CIC report from ICIC mobile app from Customer's phone

• 

  OTP screen to get customer consent to process customer data

AutoTrack AI — Loan Simulator

Inspiration

Closing a car loan often fails due to complex calculations and "hidden" credit requirements.
We built AutoTrack AI to bridge this gap, turning natural conversations into structured financial data and real-time credit assessments.

What it does

An AI-powered simulation suite that lets users talk their way through a loan application.

• Calculates 4-variable financial models
  
• Automates UI form updates via chat
  
• Analyzes CIC (Credit Information Center) PDF reports
  
• Determines eligibility instantly
  

How we built it (AI Process)

The core of our project is an Agentic Workflow built with NestJS and Qwen-plus (Alibaba DashScope):

1. Structured Output Engine

We enforced a strict JSON-only response format using complex system prompting.

Our parseAssistantPayload() layer ensures that even if the AI adds conversational text, the system extracts only valid JSON to drive the UI.

2. Stateful 5-Step Logic

The AI manages a state machine:

Welcome → Financial Info → Target → Selection → Ongoing

This guides users through a logical funnel while maintaining context across the session.

3. Real-time Field Patching

We built a bridge between the LLM and the React frontend.

When the AI identifies financial data (e.g., "I earn 30M/month"), it emits a fields object that triggers a patchForm() event, updating the UI sliders and charts in real-time.

4. CIC PDF Extraction

We implemented a pipeline that:

• Parses raw PDF text
  
• Uses the LLM to extract key metrics:
  • Total Debt
    
  • Monthly Obligations
    
  • Debt Group
    

This data is fed into a DTI (Debt-to-Income) engine to calculate eligibility.

Challenges we ran into

Prompt Robustness

Preventing the LLM from outputting Markdown code blocks or preamble text that breaks JSON parsing.

Solution:
Implemented an extractBalancedJsonObject algorithm.

VND Processing

Handling various Vietnamese currency formats:

• "30tr"
  
• "1.5 tỷ"
  

Solution:
Regex preprocessing combined with LLM reasoning to normalize all values into a standard VND integer.

API Latency

CIC analysis can take up to 60 seconds.

Solution:

• Nginx streaming configurations
  
• Proactive UI loading states
  

Accomplishments that we're proud of

• Successfully implemented AI-Driven UI Control where chat acts as the primary controller for a complex financial form
  
• Achieved high accuracy in Automated CIC Analysis, converting a 5-page report into a simple Eligible / Ineligible result within seconds
  

What we learned

• How to handle hallucinations in financial data by adding a secondary validation layer (Sanitization) between AI output and the database
  
• The effectiveness of Single Source of Truth (SSOT) to ensure both AI and calculator engines use identical banking interest rates
  

What's next

Fine-tuning

Train a local Llama 3.2 model on Vietnamese banking regulations to:

• Reduce API costs
  
• Improve data privacy
  

Omnichannel Integration

Deploy the AI agent on:

• Zalo
  
• WhatsApp
  

→ Capture leads directly from social messaging platforms

Built With

• nextjs
• node.js
• typescript