BluePrint | Devpost

America is in a financial crisis

As of early 2026, the weight of consumer debt has reached staggering heights:

Total Household Debt reached a record $18.59 trillion, driven largely by mortgage and auto loan costs.

Credit Card Debt: Total credit card balances have surpassed $1.23 trillion.

Interest Trap: With average credit card APRs hovering around 21%, many Americans are paying more in interest than they are on the actual goods they purchased.

We noticed that most people are drowning in two types of data: "messy" documents like tax returns and bank statements, and "complex" market math.

We wanted to build more than a chat bot; we wanted a partner. Inspired by the need for a truly unified system, we built Project BoostUp to bridge the gap between your personal files and the professional financial world.

What We Learned:

We started at the drawing board with architectural diagrams and schemas.

Building an agentic system taught us that accuracy is everything.

We realized that while AI is great at talking, it can sometimes struggle with precise numbers. To build trust with potential users, we had to combine the reasoning power of an AI with the exact mathematical certainty of a professional engine that could provide financial statistics. We also learned that keeping data safe is a high priority when dealing with AI, leading us to build a dedicated cleaning system that strips out private info before the AI ever sees it.

How We Built It

BluePrint is built on a fast Python backend designed to think and act on its own.

The Brain (OpenAI Agent SDK): We utilized the OpenAI Agent SDK to orchestrate our AI's actions. This allowed us to build an agent that doesn't just answer questions but actually "plans" how to solve a problem—whether that’s looking up a stock or reading a PDF.

The Precision (Wolfram Engine): When the agent needs to do real math, it hands the work over to Wolfram Alpha. This ensures that every calculation, from interest rates to portfolio risk, is 100% accurate.

The Memory (RAG Pipeline): We used a "Retrieval-Augmented Generation" pipeline. When you upload a doc, our system cleans it, breaks it into smart chunks, and stores it so the AI can "remember" and reference your specific financial history.

The Look (React & Tailwind): We designed a sleek "Midnight Emerald" dashboard that feels like a high-end financial command center, complete with real-time animations as the AI works.

Challenges We Faced

The "Hallucination" Wall: Early on, the AI would try to guess tax rules. We fixed this by forcing it to check Wolfram’s verified financial calculations instead.

Messy Documents: Standard tools often struggle with the weird layouts of bank statements. We had to create a custom "header injection" technique so the AI always knows which account it’s looking at.

Duplicate Data: To keep the system fast, we built a hashing system that recognizes if you’ve uploaded the same document twice, preventing the AI from getting confused by repeated info.

Security: Redacting documents using Microsoft Presidio to make sure that individuals didn't get their financial data like SSN's or credit card numbers scraped for AI training.

Technical Stack

Agent Logic: OpenAI Agent SDK

Math Engine: Wolfram Alpha API

Backend: Python & FastAPI

Database: ChromaDB (Vector Search)

Frontend: React & Tailwind CSS

What's Next for BluePrint?

We want to move from giving advice to taking action. Our next goal is to let the agent actually execute trades or move money into savings accounts automatically based on the smart rules you set and will constantly learn and adapt from market trends utilizing machine learning algorithms.