Building with AI agents

Kiarrow uses AI in two places you can point to: in the products, and in how the code and content are produced. Neither replaces responsibility for what is published.

AI is a feature in the products, not what they are

In StudieDoktorn a tutor sits inside the study loop. It should answer from the platform's material and allowed sources, with citations, and refuse questions about a real patient. In HPredo, generated extra practice can appear in paid plans; it is labelled as generated and is not UHR material.

In both cases the model is not the source of truth. Grading, scores, credits and access live on the server.

In the build: instructions, then checks

The HPredo and StudieDoktorn product repositories contain agent instructions: what may be published, what must have a source, and which commands must run before something is called done. That is working order, not a claim about speed or cost.

For StudieDoktorn's clinical text the order is roughly: fetch the source, write a draft against it, check the quote, then publish. A draft that cannot be evidenced should not be filled with invention.

The same rule applies to this site. Prices, packages and company details come from code and registries, not from a model guessing. Reviews, client names and figures about users or revenue are not published until they exist.

What AI does not remove

Kiarrow takes on websites, apps, dashboards, SEO and marketing. Claims about a client business are held to the same line: no invented results, no generated “proof” images, no review we do not have.

An agent can keep the structure. It cannot replace someone being accountable for the text against a source.

Summary

Building with AI agents, as we do it, means three things: the source first, a quote check where the content is sensitive, and a server that owns the answer key. It is slower than letting a model finish the page. It is a deliberate choice.

Read more about the company, HPredo and StudieDoktorn.