We initially looked at a mobile laser tag concept, but looking at the FNB challenge, we realized that most banking apps just show you the raw numbers, but they don't give you the tools to understand how to make good financial decisions. We wanted a co-pilot that helps users navigate their daily spending limits and highly important decisions, like dodging predatory loans.
What it does
LoanGuard AI is a two part financial resource.
AI loan auditor: Users can upload text-based PDF loan contracts, supporting files up to 12mb and 80 pages. The app extracts the principal, stated APR, and term, and then using a custom prompt with OpenAI gpt-4o-mini to scan for predatory clauses. It outputs a transparent risk score from 0-100 and highlights the exact problematic clauses on the original PDF.
Safe-to-spend & Loan Finder: Instead of showing only a raw balance, the app calculate a user's actual cash flow capacity by analyzing the users transaction history from a CSV upload, protecting their saving goals, upcoming bills, and a emergency fund. The loan finder checks if the amortization cost of a specific loan fits into their remaining budget.
How we built it
We built the full-stack web application in Python, using Streamlit for a reactive, bank deployable frontend dashboard. Document parsing is powered by PyMuPDF, and reportlab handles the generation of PDFs. To insure trust, we strictly separated LLM text extraction from the underlying math. OpenAI gpt-4o-mini is constrained to output JSON and exact substrings to hallucination. The model never calculates the numbers. All financial modeling is handled by deterministic Python functions.
Challenges we ran into
Our first issue was simply choosing a concept, as we spent time debating between this banking suite or a Mobile QR code laser tag app. Additionally, since we were building something that is purely web-based, it was difficult to coordinate 5 members to use the same libraries while integrating all of their projects together at the end.
Accomplishments that we're proud of
We proved our core thesis, which is that: Trust is a feature. By separating the AI's compression from the deterministic math calculators, and forcing the app to cite it sources and highlight exact contract clauses, we built something that a bank could legitimately deploy without having to worry about AI hallucinations
What we learned
We learned the realities of rapid full stack development and the importance of strict constrains when dealing with AI in finance. We discovered that you cant just let an LLM act as a chatbot regarding a legal contract, you have to engineer the prompt to force it into an objective, data-extraction format.
What's next for Team LARP FNB
Plan on polishing the app tomorrow, changing some aspects of some features.
Log in or sign up for Devpost to join the conversation.