DecisionTwin — See the Decision Before You Make It

Inspiration

A car loan can look affordable on paper and still be the wrong decision for your financial situation.

When people shop for a car, the conversation often revolves around one number: the monthly payment.

But that number doesn't tell the whole story.

What happens to your emergency savings after the down payment? How much are you actually paying in interest and fees? What if your income drops temporarily? And how do you know whether the financial terms in front of you are correct?

We realized that the problem isn't a lack of financial information. It's that people are expected to turn scattered, complicated, and sometimes conflicting information into a major financial decision.

That's why we created DecisionTwin.

Our goal is to help people understand the financial consequences of a decision before committing to it—and see the evidence behind every important number.

What It Does

DecisionTwin is an AI-assisted financial decision intelligence platform, starting with automobile loans.

Instead of simply calculating a monthly payment or generating a generic recommendation, DecisionTwin builds a personalized comparison of a user's financial position before and after a proposed financial commitment.

The experience is designed around six capabilities:

1. Intelligent Financial Intake

A knowledge graph identifies which financial details are necessary for the user's analysis. It prioritizes important missing information, explains why questions matter, and distinguishes known facts from estimates and unknowns.

2. Loan Document Intelligence

Users can enter loan terms manually. Our document intelligence module is also designed to extract terms from supported loan PDFs, cross-check the results, and present evidence for user verification.

3. Real Financial Impact

A deterministic financial engine calculates loan payments, interest, finance charges, savings impacts, and changes to monthly cash flow.

4. Financial Stress Testing

Illustrative scenarios help users explore how unexpected expenses or reduced income could affect their finances with and without the proposed loan.

5. Alternative Decisions

DecisionTwin explores feasible changes, such as different down payments, loan terms, or optional add-ons, while showing the trade-offs involved.

6. Explainable Results

Important calculations and conclusions are connected to their underlying inputs, formulas, documents, or assumptions.

The goal is not to replace the user's judgment. It is to give them better evidence for making their own decision.

How We Built It

We designed DecisionTwin as a modular financial intelligence application.

Our technology stack combines:

Next.js, TypeScript, Tailwind CSS, shadcn/ui, and Recharts for the interactive interface.
Python, FastAPI, Pydantic, and SQLite for backend services and structured financial data.
NetworkX for financial dependency modeling and adaptive questioning.
Python Decimal for precise financial calculations.
PyMuPDF for extracting information from supported PDF contracts.
Google Gemini integration for structured document understanding where enabled, with deterministic verification and privacy controls.

One important design decision was to separate AI interpretation from mathematical computation.

AI can help interpret financial language, but the actual calculations must be performed by deterministic, testable functions.

We also used a multi-agent engineering workflow with specialized development responsibilities and an independent auditing process to identify implementation errors before integration.

Challenges We Faced

One of our biggest challenges was ensuring that the financial analysis was genuinely correct.

For example, a loan's principal, amount financed, contract interest rate, and disclosed APR are not interchangeable. Treating them as equivalent can produce misleading results.

We also had to account for trade-in equity, cash down payments, prepaid finance charges, and the difference between missing information and a genuine zero.

Our independent audit exposed several issues in the initial implementation, including insufficient input validation and edge cases in financial questioning.

We addressed these through explicit financial contracts, regression tests, and correction directives.

Another challenge was making a complex financial analysis understandable to someone without a finance background.

What We Learned

We learned that building trustworthy financial AI requires more than a capable language model.

A convincing answer is not necessarily a correct answer.

Financial calculations need independent verification. Extracted document values need evidence. Missing information must remain visible. And users need to understand which results are facts, calculations, or hypothetical scenarios.

We also learned how important continuous integration and independent auditing are when multiple engineering components are being developed concurrently.

What's Next

Automobile financing is our first use case, but the broader vision is to create a financial decision intelligence platform that can support other significant commitments.

Future improvements include broader document support, more financial products, lender-offer comparisons, and additional scenario-planning capabilities.

We want DecisionTwin to make one thing possible: before making an important financial decision, people should be able to see its consequences, understand the evidence, and decide with confidence.