Building useful software
Why Car Deal Check keeps the math outside AI
A dealer worksheet mixes several kinds of information: the vehicle price, optional products, taxes, trade-in figures, and financing terms. Reading it is one problem. Working out how those numbers fit together is another. Car Deal Check treats those as separate jobs.
When AI import is available, a model can turn a pasted quote or a photo into structured fields. That is a useful role for a language model: recognizing that a line labeled “cash down” belongs in the down-payment field, for example. It is also a place where mistakes can happen. An unclear scan, an unusual label, or two similar totals can produce an incorrect reading.
The application checks the returned structure and converts the amounts into cents before using them. Missing values remain missing. The extracted information becomes editable, so a person can compare it with the original worksheet and correct it. A plausible-looking answer is not enough to accept an amount without a way to inspect it.
Once the inputs are known, ordinary calculation code handles the numbers. Trade equity is the allowance minus the payoff. For a fixed-rate estimate, the amount financed, rate, and number of payments determine the monthly payment. Changing the same input should produce the same result each time; it should not depend on how a model phrases its response.
The distinction also matters for explanations. A generated summary can suggest questions about an optional product or an unexplained fee. It does not get to replace the calculator’s financial results with different numbers. If an AI explanation is unavailable, the application can still show the calculation and questions assembled from the entered information.
This approach makes a useful manual mode possible. You can enter the figures directly, compare rates, and explore trade equity without uploading a document or making an AI request. The public calculator currently works this way while AI import is paused. A model outage or a usage limit should not remove the underlying utility.
There are limits to what a calculator can establish. It cannot confirm that a dealer entered the right tax, determine whether a particular charge is lawful in every jurisdiction, or turn a payment estimate into a lending offer. Rounding, payment timing, and the actual contract can also affect the result. The interface needs to make those boundaries visible alongside the answers.
For a small product, this is a practical design habit: identify the part that benefits from interpretation, keep the rules you can express explicitly in code, and let the user inspect the handoff between them. AI is most useful here when it saves some typing while leaving the important decisions understandable.