Built with AI tools, shipped with no AI features

How the app was made

I’ll say this plainly because it’s already public and because hiding it would be silly. I didn’t write kisir’s code by hand. I described what each screen should show and what each rule should be, an AI coding tool wrote the server, the database layer and the app screens, and I checked the result on my phone and said what was wrong. Then it fixed it, and I checked again. That loop, many hundreds of times, is how the app got built, in Armenian, Russian and English, for iPhone and Android.

The same goes for a lot of the words on this site. Drafts get help from the same kind of tool, and I read every line before it goes up, because a draft that’s wrong about the app is worse than no draft.

So kisir is, by any honest description, an app made with AI.

What’s in the app

Now the other fact. Open kisir and look for the AI. There isn’t any.

When you add an expense, the app divides, in whole cents. When you open the overview, it adds. When you settle up, it runs a fixed procedure that finds the fewest transfers, and the settle-up guide lets you reproduce the answer by hand. It doesn’t scan receipts, sort your expenses into categories, predict your split, write summaries, or time its notifications. For an expense in a foreign currency it fetches today’s exchange rate, which is a table lookup you can overwrite.

The app you install contains no model. Nothing in it learns from you. The longer argument is in why kisir has no AI inside.

Why those aren’t in tension

A tool used to make a thing and a feature inside the thing are different in exactly the way that matters.

When the coding tool writes the settlement procedure, I can read the procedure. I can test it with four friends and a villa and see that the balances sum to zero. If it’s wrong, it’s wrong in a way I can see, fix, and test again. By the time it ships, the code is a fixed set of rules that does the same thing every time. The tool’s uncertainty was resolved before anyone’s money went through it.

A model inside the app is the opposite. Its uncertainty is resolved at the moment you use it, on your expense, with your friends watching. If a receipt scanner misreads a 7 as a 1, that’s in the ledger now. If a “smart split” guesses that Ben had the steak, Ben is paying for a steak. There’s nobody checking first, because the whole point of the feature is to skip the check.

So the rule I ended up with is simple. Use any tool you like to build it. Ship only what the group can verify. An AI coding assistant passes that rule, because its output is code that gets tested. An AI feature in a bill-splitting app fails it, because its output is a number that gets trusted.

What the tool was good for

Since I’m being honest about using it, I should be honest about what it did well, because it’s not what people assume.

It was very good at the parts where the rule was clear and the typing was long: three languages of interface text with their own plural rules, a form that refuses to save until the amounts add up, the export to CSV and Excel, a history line on every edit. Those are features that are tedious to write and easy to specify, and they’re most of an app like this.

It was not where the decisions came from. Nobody asked the tool whether the app should have accounts; the answer was no because of how friends actually join a group chat. Nobody asked it whether the dram should settle in whole units; that came from a real group being told to pay 0.99 of a dram, which is in the post nobody can pay 0.99 of a dram. Nobody asked it whether to scan receipts. Those are product decisions, and they came from using the app on real trips and flats, and from the people who complained.

Why the product stays plain

In 2026, “made with AI” is on the box of almost everything, and “has AI inside” is the feature list almost everyone is adding. kisir’s answer is to be exactly the kind of software that doesn’t need it.

A split expenses app has one job: four friends, one list, one answer they all agree on. The answer has to be the same on every phone and reproducible on a napkin. Every piece of that is arithmetic and rules. Adding a model to it doesn’t make it more capable. It makes it less checkable, which is the only quality the thing has to have.

So the tools stay in the workshop. What ships is a form, a list, an overview and a plan, in 20 languages, free, with no account and no ads. If you want to see how it compares to apps that chose the other way, the best apps to split expenses guide is honest about both. And if you just want to try the plain one, the no-sign-up guide shows how a friend joins in ten seconds, which is the feature that actually matters.