Skip to main content Skip to footer

From pilot to air traffic controller: what it takes to rebuild an app with AI

For around ten years, a small mobile app has been helping doctors in some of the hardest places in the world get advice from specialists they'd otherwise never reach. This year, the Microsoft technology it was built on came to the end of its life, and the app needed rebuilding from the ground up.

A few years ago, that would have meant a long, painful project. This time, one of our developers rebuilt it in a relatively short time, with AI tools writing the code and an experienced developer steering every step.

If you run a circular business, you've probably heard plenty about AI building apps in an afternoon, and you may have tried it yourself. This is what it looks like on a real, complex app that people depend on, and why the expertise around the tools matters as much as the tools themselves.

Bringing a specialist into the room

Collegium Telemedicus, or CT, connects doctors working in the field with volunteer medical specialists. Its focus is on regions where specialist care is extremely difficult to access, places like Gaza and Afghanistan.

The idea is simple and powerful. A doctor creates a case in the app, adding as much detail as they can. A volunteer specialist reviews it and advises on a treatment path, or at the very least on ways to improve the patient's situation. It bridges the gap between generalist doctors on the ground and the specialists who can help from a distance.

CT began life as a basic website and app built by another developer. Barry, our founder, later moved it into a proper mobile app, and several of our team have worked on it over the years. It also has a sister app, Save a Child, which shares the same foundations with its own styling and a few interface changes.

When the technology underneath runs out of road

The original app ran on a Microsoft mobile technology, Xamarin. Its code history goes back to December 2016, though Filipe, the developer who led the rewrite, suspects some earlier work isn't captured there. Either way, it had been doing its job for the best part of a decade. Then Microsoft retired the technology. That meant no more security patches and no support for new features. For an app that holds details of real patients, carrying on without security updates wasn't a comfortable place to be.

The code itself had also become brittle. A change in one part of the app would frequently break something somewhere else, and there was no automatic way of catching it. Across the entire app, there were only around 20 automated tests. These are small checks that run on their own and confirm each piece of the app still does what it should. With so few of them, the only way to be sure a change hadn't broken something was for a person to click through the whole app by hand.

Rebuilt from scratch, now with 640 safety checks

Filipe hadn't worked on the original app, and he rebuilt it entirely using AI tools. The new version runs on React Native, a framework that's actively maintained and, as Filipe points out, among the three most widely used for building mobile apps. Like the original, it works on both Android and iPhone, but keeping the two reliable is far more manageable now. Filipe expects it to have a much longer supported life than the technology it replaced.

The app does everything it did before, with bugs fixed along the way. As Barry puts it, before AI this would have been a really painful project. Filipe completed it in a relatively short time, using AI throughout to write the code, produce the documentation, and create and run the automated tests.

The biggest difference shows up in those automated tests, where the app went from around 20 of them to over 640. Each one confirms that a piece of the app's logic keeps producing the result it should. If a future change breaks something, the tests catch it straight away, rather than someone stumbling across the problem later. For an app used in such difficult places, that's a very different footing to build on.

The testing has kept growing since the rebuild, too.

Filipe recently added a new kind of check that tests the whole app from start to finish, using a new feature in Expo, the tool that builds the app and ships it to the app stores. It launches the app in a simulator, taps around following the journeys real users would take, and grabs screenshots along the way. That alone is already helping to catch bugs, and when the screenshots are passed to AI, it can pick out a few more issues by itself.

From pilot to air traffic controller

When Filipe talks about how his work has changed in recent years, he describes it as going from being a pilot to being an air traffic controller. As a pilot, he had his hands on the controls the whole time, writing every line of code himself. Now his job is to set the direction. He gives the AI context about what's needed and why, then reviews the plan it proposes for how to do the work. Once he's happy with the plan and approves it, the AI carries it out, tests included.

He's committed enough to this way of working that he actively avoids typing code himself. When he does, he's found the AI sometimes ignores what he's typed and comes back with questions instead. That shift puts experience at the centre of the job rather than at the edges. Knowing what a good plan looks like, spotting when one is heading the wrong way, and deciding what to approve all depend on years of building software the hard way.

Drawing the house before building it

Filipe has a simple way of explaining why AI needs direction. Ask it to draw a house and you could get almost anything back. Ask for a house with two floors, two bedrooms and a roof, painted blue or white, and what comes back will be much closer to what you had in mind.

Over time, we've built that thinking into how every project runs. Each one has a folder of development documents that sits alongside the code, holding the plans and records the AI works from. Before any development begins, Filipe and the AI put together a plan for the work. He reviews and approves it, and it's saved with the code so the AI can keep referring back to it as it goes.

Whenever the app needs to rely on a new piece of outside code, a decision record has to be written first. It notes who approved it, when and why, and it stays with the project permanently. That leaves a clear trail of how the app came to be the way it is. The more of these rules and constraints are written down, the more focused and predictable the AI's work becomes. They also stop the AI from undoing work that's already been finished, which was a real problem before these documentation habits were tightened up.

What this means if you're building something circular

AI tools have made it possible to move remarkably fast, and we use them every day on serious work. That speed is only part of what made this project succeed.

A decade-old app that real people rely on was rebuilt comparatively quickly, and it came out with far stronger safety checks, clear records of every decision, and a foundation that should last for years. That came from pairing the tools with experienced developers who know how to plan the work, keep the AI on course and judge what it produces.

=====

If your own platform is sitting on ageing technology, or you've taken an idea as far as AI tools alone can carry it, we'd be glad to talk it through with you.

About the author

Paul Smith

Paul is a self-described communications Swiss Army Knife, having spent the majority of his career supporting beneficially impactful companies across the spectrum effectively tell their story to the world. The circular economy is a long time love of his, so he’s thrilled to now play a part in expanding its impact in the world. When not behind the computer, Paul can usually be found behind a book, on his bike exploring, or out for a walk in the forests of Fontainebleau, getting a closer look at the latest moss…