AICalc

The Task: Accuracy and Usefulness

We wanted to create a service that provides quick and accurate answers to practical questions, from calculating materials for renovations to financial computations. The main requirement for such a product is trust. The user must be confident that the result is correct and the methodology is transparent. Therefore, we decided from the outset that every formula would be checked against a primary source, and the calculation modules themselves would be isolated and covered by tests.

In addition to the calculations themselves, it was important to provide context: how to prepare the data correctly, what mistakes to avoid, and how to interpret the result. This transforms a simple tool into a useful advisor. This led to the idea of supplementing each of the 82 calculators with an in-depth analysis. At the same time, it was necessary to ensure convenient output of results, including printable forms and consolidated estimates for complex tasks like renovations.

Engineering Solutions

The site is built on a static architecture using Astro with Tailwind, hosted on Cloudflare Pages. This ensures high loading speeds, as the pages are served pre-rendered. For dynamic data, such as currency exchange rates, we use Cloudflare serverless functions.

The core element is the calculation engine. The formulas are placed in separate modules, independent of the interface. They are checked by 867 automated tests: a change that breaks a formula does not go unnoticed. A single template was developed for printing results, which generates an A4 sheet with all the necessary data, including the methodology and a QR code.

For client projects, this means we know how to build reliable calculation services. We separate logic from presentation, cover it with tests, and design an architecture that combines the speed of a static site with the flexibility of dynamic functions. This is important for any product where the cost of an error is high.

Controlled Content Automation

To fill the 82 calculator pages with useful analyses, we used generative AI. The Gemini model, via Vertex AI, received only the page text and the code of the corresponding calculation module for analysis. The model does not see other sources, so its every statement can be verified against the page. We build description generation for client tasks in the same way: first, a list of facts, then the text.

Generated text cannot be trusted without verification. That's why we developed a three-layer control system. First, an automated process searches the generated text for numbers or names that are not in the source data. Then, a separate call to a reviewer model looks for false statements, and the model corrects what it finds up to two times. Text that does not pass these filters is not published until reviewed by a human. Out of 82 analyses, 50 passed the checks automatically, while 32 required manual edits.

This experience shows how we work with AI: not as a magic wand, but as a tool that requires supervision. Our verification system not only improves content quality but also helps find flaws in the product itself — for instance, the reviewer incidentally found 3 flaws in the calculators themselves, and we fixed them. For your project, this means we implement automation responsibly, with an established quality control process.