With FlyHere, AI builds more than features. It builds the maintenance too
The drone zone map refreshes itself, checks itself and reports to my phone.
FlyHere is my app for drone pilots. It shows where you may fly and where you may not. I build it alone and AI agents write most of the code. I keep improving it.
The whole product stands on one thing: the map of zones where flying is banned or restricted. Aviation authorities publish that data and it changes. When it's stale, the app lies. And with a drone, a lie means a fine or something worse.
I had the features. The problem was elsewhere
The agents built me the map, the "can I fly here" verdict, an AI assistant, the web, iOS and Android. Features piled up fast. That's the part people talk about at conferences.
But I refreshed the zone data by hand. A new cycle from the authority meant a rebuild, approving the publication and a web deploy. A cycle has a date when it expires. It was the only dated thing in the whole project. And that very thing depended on me remembering it.
In September, one web deploy took down the hosted zone data and the map. The second trap was worse. The phone app didn't download the data and quietly fell back to the version it was built with. No error, no log. The map looked fine and it wasn't.
One sentence of instructions
On October 23 I gave the agents one sentence: I want this to happen automatically, and only statuses should come to me in Telegram.
The whole machinery was born that day. It runs by itself and downloads the data straight from the authorities. It checks the Czech zones several times a day, other countries daily, and a summary arrives in the morning.
The next day seven countries were added. Today sixteen run on a single data address. Web, iOS and Android read exactly the same data. A web deploy can't break them anymore, because the data lives separately.
Publishing is just switching a pointer to the new version. The old version stays stored next to it. When something goes wrong, going back is again just a switch. No new build, no new release on three platforms.
Telegram doesn't get every run and every "finished OK". Nobody would read that after a week, not even me. Only what means something comes through.
Messages on my phone over one made-up week. Countries XY and XZ and the numbers are made up.
- Monday: Morning summary: all layers fine.
- Tuesday: CZ zones: new version out, 3 changes.
- Wednesday: XY zones: the license changed, holding the last good version.
- Thursday: XZ zones: only 61% of zones arrived, nothing goes out.
- Friday: Morning summary: one country holds the last good version, the others are fine.
Five messages a week. The rest runs quietly.
Refreshing is easy. The checks are the work
Any script can download a file and upload it somewhere else. What matters is what happens when something goes wrong. This is where the agents did the most work.
Each check has its reason:
- License. I approved each country once and the system stored a fingerprint of the license text. When the license changes, the country holds the last good data and I get an alert.
- Shrinkage. When a new set arrives with fewer than 85% of the zones, nothing goes out. It's more likely a download error than an authority that opened the sky overnight.
- Answers. For countries where the app gives a verdict, the system runs the engine on checkpoints at every refresh and compares the result with signed answers. Doesn't match? The layer stays display only, one alert arrives and it tries again the next day.
- Silence. A layer that hasn't refreshed for a long time reports itself.
- Brake. Any country can be pulled from the app with one command, in about two minutes. The rollback holds the stop until I knowingly release it.
We hit a limit on the very first day. The next day the agent added a budget, an early warning and the order in which countries get priority.
When the authority changes the data
On October 29 the Czech authority withdrew dozens of sectors in one area. The check failed. Exactly as it should. Without it the change would simply have gone through and nobody would have known anything happened.
It turned out the test was too narrow, tied to specific sectors. The agent rewrote it to guard the whole data set. I found out from my phone, not from an angry user.
An agent builds a button and a license check with equal willingness. The button is visible, the check isn't.
Why this decides things in a one-person project
In a company, someone is responsible for maintenance. In a side project, nobody is. I have emable, clients, courses. FlyHere gets time when some is left.
If I used agents only for features, every month I'd have more features on top of ever older data. A quickly built product that quietly rots. And the worst part is you can't tell from the outside. The map lights up, the verdict answers, it's just no longer valid.
So I have to assign maintenance and checks explicitly. They won't appear on their own.
Four questions for your data
If you build anything with AI that stands on outside data, ask yourself four questions. It doesn't have to be drones. A supplier's price list, exchange rates, a list of branches, a legal text. Anything someone else changes for you.
- What in the product ages on its own, and who refreshes it today?
- How do I tell that the new data is worse than the old?
- What happens when a refresh quietly doesn't run?
- How do I roll it back with one command?
Where the answer is "me, by hand", give the agent a task. Not a feature. Machinery. It looks boring and nobody will praise it. But thanks to it the zone map in FlyHere refreshes even in a week when I don't touch the project.
You can find FlyHere at flyhere.app.