Inspiration We wanted to tackle the frustrating experience of applying for loans. Traditional loan applications feel like throwing paperwork into a void and hoping for the best. There's no transparency, feedback takes forever, and applicants rarely understand why they were approved or rejected. We thought, what if we could make the process instant, explainable, and actually helpful? That's where LoanAI came from. The goal was to build something that gives people real-time feedback on their applications while being upfront about how decisions are made.
What it does LoanAI is a desktop application that evaluates loan applications instantly using AI. Users fill out a form with their personal, employment, and financial information. As they type, the system validates their phone numbers, emails, and addresses against real external services. Once submitted, the application goes through an AI-powered evaluation that considers data quality, financial health metrics like debt-to-income ratio and credit score, and overall risk factors. Within seconds, users get a decision along with detailed reasoning, a risk assessment, and suggested loan terms if approved. The whole point is transparency. Users actually understand why they got the result they did.
How we built it The frontend is an Electron app using plain HTML, CSS, and JavaScript. We used Three.js to add some visual polish to the login screen and GSAP for smooth animations throughout. Firebase handles user authentication.
The backend runs on Node.js with Express. It connects to Groq's API for the AI evaluation, Abstract API for email validation, and OpenStreetMap's Nominatim service for address verification. The AI model uses a carefully crafted system prompt that balances thoroughness with fairness, making sure it catches fraud indicators without being unnecessarily harsh on legitimate applicants.
We kept the design minimal and professional. Blue color scheme, clean typography with Cormorant Garamond and Inter, no flashy gradients or over-the-top animations. We wanted it to feel like a real financial application, not a tech demo.
Challenges we ran into Getting the AI evaluation to be balanced was tricky. Early versions were either too lenient (approving obviously fake applications) or too strict (rejecting reasonable applicants for minor issues). We spent a lot of time refining the system prompt to strike the right balance between fraud detection and fair assessment.
Real-time validation was another challenge. Coordinating multiple external API calls without making the UI feel sluggish required careful async handling. Rate limits on free services like Nominatim also meant we had to be smart about when and how often we validated addresses.
Packaging the Electron app for distribution had its own headaches. Making sure the backend server starts automatically, handling environment variables properly across different machines, and dealing with Windows Defender flagging unsigned executables all took more time than expected.
Accomplishments that we're proud of The instant feedback loop is something we're really happy with. Users don't have to wait days or weeks to know if their application looks good. They get actual, actionable information immediately.
The fraud detection is genuinely useful. The AI catches placeholder data, inconsistent information, and unrealistic financial claims without being a black box about it. When it flags something, it explains why.
We also managed to build a complete, polished application that actually works end-to-end. It's not just a prototype or a mockup. You can log in, fill out an application, get a real AI evaluation, and understand the decision. That feels good for a hackathon project.
What we learned We learned that prompt engineering for financial applications is harder than it sounds. You need to balance multiple competing concerns: catching fraud, being fair to applicants, explaining decisions clearly, and handling edge cases gracefully_. It took many iterations to get right.
We also learned the value of real validation. Mocking everything is faster for demos, but integrating with actual external services exposed issues we never would have found otherwise. Real-world APIs fail, return unexpected data, and have rate limits. Building with that in mind from the start made the final product much more robust.
On the technical side, Electron's quirks around packaging and distribution were a learning curve. Understanding how to bundle a backend server with a frontend app in a way that works reliably on other machines taught us a lot about deployment considerations.
What's next for LoanAI We want to add more sophisticated document verification. Right now we validate contact information, but imagine being able to upload pay stubs or bank statements and have the AI verify them against the claimed financial information.
A mobile version would make sense too. Most people apply for loans from their phones these days, so wrapping this in a React Native or Flutter app could expand the reach significantly.
We're also interested in building a lender dashboard. Right now it's focused on the applicant side, but lenders need tools too. Showing aggregate risk metrics, approval trends, and fraud patterns would make this useful for both sides of the equation.
Finally, we'd like to explore fine-tuning the AI on actual loan decision data. The current model uses general reasoning, but training on historical approvals and rejections could make the assessments even more accurate and aligned with real-world lending standards.
Log in or sign up for Devpost to join the conversation.