YOUSUN Amicus: AI Agent for Fair Lending

YOUSUN Amicus goes further.

Inspiration

YOUSUN Amicus started from a simple but very real question: what happens when a good small business needs financial support, but the bank cannot clearly understand its real repayment capacity?

Many small businesses do not have a perfect financial profile. Their income can be seasonal, their cashflow can change every month, and their documents may not always tell the full story. At the same time, bank officers also face pressure. They need to review applications quickly, reduce risk, and still make fair decisions.

In many lending systems, the journey ends with a simple approval or rejection. But I wanted to build something more responsible than that.

The inspiration behind YOUSUN Amicus is this belief:

Good lending should not only approve loans. Good lending should understand risk, explain decisions, and support customers before default.

That is why the project focuses on three ideas:

Simulate before lending. Report with fairness. Rescue before default.

What it does

YOUSUN Amicus is a two-sided AI agent platform for responsible SME lending.

It serves both the customer and the bank officer.

On the customer side, a borrower can log in, submit a loan application, provide business details, select available documents, and track the application status. The system also guides the customer by explaining why documents like trade license, bank statement, or supplier invoice are important.

On the bank side, a bank officer can log in, view applications, open an application detail page, and run the Amicus agent workflow.

The agent can:

Build a Financial Twin of the borrower
Simulate repayment risk through Future Mirror
Recommend a safer loan amount using Debt-Trap Shield
Suggest realistic repayment guidance with Dynamic EMI Planner
Generate Gemini-powered reports
Save reports and analysis to MongoDB
Create a Rescue Before Default plan before the borrower reaches default

This makes YOUSUN Amicus more than a chatbot. It does not only answer questions. It performs real workflow actions: submit, analyze, report, save, and rescue.

The main agent workflow is:

Observe → Understand → Simulate → Recommend → Report → Rescue

How we built it

We built YOUSUN Amicus as a full-stack web application with a live deployed demo.

The frontend is built with Next.js, TypeScript, and Tailwind CSS. It includes a professional bank officer dashboard, customer portal, application pages, report pages, and rescue workflow UI.

The backend exposes API routes that work as agent tools:

runLoanReviewAgent
generateGeminiReport
createRescuePlan

MongoDB Atlas is used as the system of record. It stores:

loan_applications
agent_analyses
reports
rescue_plans

Gemini is used to generate AI-powered reports such as bank officer memo, customer-friendly summary, and rescue-before-default report.

The project also includes:

OpenAPI-defined tool specification
Agent Builder-ready system prompt and workflow documentation
MongoDB MCP Server configuration example
MongoDB MCP tool mapping
Public GitHub repository
Live Vercel deployment
Production API proof

The production APIs have been tested successfully. The loan review API returns success and saves analysis to MongoDB. The Gemini report API generates reports, and the rescue plan API creates early support plans.

Challenges we ran into

The biggest challenge was turning the idea into a real agent workflow instead of just a beautiful dashboard.

At first, it was easy to imagine the UI, but the harder part was making the system actually perform actions. We needed the customer application to be saved, the bank officer to review it, the agent to analyze it, the report to be generated, and the rescue plan to be stored.

Another challenge was making the project feel useful for both sides. A bank-only dashboard would not be enough. So we added a customer portal where a borrower can submit an application and understand the process more clearly.

We also faced deployment and cloud setup challenges. Google Cloud billing was not available during the final deployment process, so the live demo is hosted on Vercel. However, the repository includes Agent Builder-ready OpenAPI tools, system prompt, workflow documentation, and MongoDB MCP proof configuration for judging and testing.

The important thing was to stay honest and still deliver a working product.

Accomplishments that we're proud of

We are proud that YOUSUN Amicus became a working deployed MVP, not just an idea.

The project now includes:

A live deployed application
A professional landing page
Bank officer login
Customer login
Customer application submission
Application status tracking
Bank dashboard
Application review page
Financial Twin analysis
Future Mirror scenarios
Debt-Trap Shield recommendation
Gemini-powered report generation
PDF report download
Rescue Before Default workflow
MongoDB Atlas data persistence
MongoDB MCP proof files
Agent Builder-ready OpenAPI tools
Public GitHub repository with README and license

The strongest accomplishment is that the production APIs are working and saving data successfully.

The project proves that YOUSUN Amicus can run a full lending workflow from customer submission to bank review, AI analysis, report generation, and early rescue action.

What we learned

This project taught us that building an agent is different from building a chatbot.

A real agent needs tools, memory, workflow, and a clear action path. It should be able to use data, make structured decisions, call APIs, save results, and support users through a process.

We also learned the importance of persistent state. MongoDB became the memory layer of the system. It stores applications, analysis, reports, and rescue plans so the agent workflow does not disappear after one interaction.

Another major lesson was responsible AI. In financial services, AI should not replace human decision-makers. It should support them. That is why YOUSUN Amicus is designed as a decision-support system. Final loan approval, restructuring, or recovery decisions remain with authorized bank officers.

We also learned that a strong hackathon project is not only about code. It also needs a clear story, honest limitations, proof screenshots, documentation, and a demo flow that judges can understand quickly.