My Bots Built a Website in Four Days. Nobody Kept the Blueprints.

I asked my team two simple questions. How much code is behind ultimatepassportguide.com? And where's the record of every change they made along the way?

The first answer came back fast: 7,229 lines, all written by the bots. The second answer was silence. There was no record. Nobody had kept one.

They had built a site covering 199 passports in 6 languages, put new versions live nine times in 48 minutes on the morning of 16 October, and never once saved a version history. It's like walking up to a finished bridge, watching the traffic flow, and learning that nobody kept the blueprints.

That's the most honest place to start.

What they built

The site answers one question every traveler has: with my passport, what do I need to get into that country? A visa, an online visa, a travel authorization, a visa at the airport, or nothing at all.

The numbers, as of this morning:

• 199 passports and 6 languages, which comes to 39,402 passport-and-country combinations. The answers match in every language.
• 27,894 individual entry rules. My team confirmed 19,991 of them by reading the government's own website. 7,609 are still being checked, and 294 haven't been looked at yet.
• 7,229 lines of code, every one written by the bots.
• Quality control: a separate bot spot-checked 236 answers against known-correct ones and found no serious mistakes. In a random audit of 2% of the data, it has finished 281 of 400 checks so far, with zero errors.

Every answer on the site shows where it came from and the date someone read it.

The cast: my AI chief of staff set priorities and kept everyone moving. A web-builder bot made the site. A research bot, Carolina, studied the competition. Perry, our AI legal strategist (not a lawyer, and he tells you so first), wrote the rules nobody was allowed to break. A quality-control bot spent its days trying to prove the others wrong. And my AI finance chief went looking for receipts.

The hard decision

The first version was fast and good-looking, and it sat on the wrong foundation: data copied wholesale from a commercial passport ranking.

Perry was blunt. A fact like "Mexicans need a visa for country X" belongs to nobody. Copying someone else's entire database is a different story, and it's the most exposed position you can take. There was also a simpler problem. If you didn't read the source yourself, you can't stand behind the answer on the day a traveler gets turned away at the gate.

So we threw it out. Every rule had to come from the destination country's own government: its foreign ministry, its immigration office, its embassy, or European Union law. The system airlines use to check passengers is a useful second opinion, but it never gets the final word.

That made the project slower and a lot harder. In exchange, every answer on the site comes with a receipt.

What broke

Plenty.

The quality-control bot flagged a batch of answers for Peru as wrong. It turned out the mistake was its own: it had been reading a list meant only for diplomats. The checker needed checking.

One government page cited for the United Arab Emirates had no visa information on it. What we'd saved was an error page. Those answers went back to "being checked" until someone finds a real source.

And when two official sources disagreed, the bots' instinct was to quietly pick one. That's exactly the kind of call a machine shouldn't make on its own. Perry's rule: show the traveler the stricter requirement, cite both sources, and mark it as being checked. A human settles the conflict.

That's why 7,609 rules still say "being checked." It isn't a failure. It's the system being honest about what it doesn't know yet. Every correction goes into a public log.

What it cost

Here's the full bill, from my finance chief's review on 16 October:

• Hosting: $0. The site runs on Cloudflare's free plan. At launch we move to their paid plan at US$5 a month. I've approved it.
• Domain names: ultimatepassportguide.com and ultimatevisaguide.com, registered with IONOS on 15 October: $21.00 total, GST included.
• AI and outside services: $0 so far.

So the whole site, as it stands, cost about 21 dollars, plus 5 a month once it launches.

The calendar: the first file showed up on 15 October at 10:30 p.m. Pacific, and launch is set for midnight on 20 October. That's 97.5 hours. Nobody logged the bots' working hours, because nobody has to pay them overtime.

And with a human team?

This is my estimate, not a quote. Here's the math, so you can argue with it:

• Research: 27,894 rules at one minute each, to read the government page, write it down, and note the source and date. That's about 465 hours.
• Programming: 7,229 lines at 50 tested lines per developer per day, which is my assumption. That's about 1,157 hours.
• Total: around 1,622 hours, or 40 person-weeks. At a blended rate of US$75 an hour, that's roughly US$122,000.

That leaves out translation into 6 languages, design, legal review and audits, so the real number would likely be higher.

Roughly US$122,000 and most of a year for a human team, against about 21 dollars and four days for mine. I'm not claiming the bots are worth a hundred thousand dollars of engineers. I'm saying the price of a first version just fell through the floor, and the hard part has moved somewhere else.

The principle

George Westinghouse designed the railway air brake backwards from what you'd expect. If the line loses pressure, the train doesn't lose its brakes. It stops. When something fails, the system fails safe.

That's what I want from an AI team. The bots will make mistakes. What matters is which way they fail. When sources disagree, show the stricter rule. When something isn't confirmed, say so. And the next time someone asks for the blueprints, I want a better answer than silence.

Speed impresses people. Knowing which way your system fails is what lets you sleep.

---

The site: ultimatepassportguide.com

If you want to learn how I run my AI team: fabiandelafuente.com/university/