18 Months Ago I Failed to Build an App. Now I Ship Them Solo With AI.

18 months ago I tried to build a cycling coaching app and gave up. This week I shipped an app that scores sales calls in seconds, grades reps against a rubric, and emails them a coaching report, all without a team, a CTO, or a single investor.

Honestly, the main thing that changed was the AI, not me. Here is the story, and what I think it means for anyone sitting on an idea they have not built yet.

The failure

Early last year I started building a cycling coaching app. The idea was straightforward: you input your heart rate zones and power data, the app generates a customised training plan around your physiology. I got it 95% of the way there.

Then I hit a set of bugs I couldn't shift. Tricky state management issues, edge cases in the zone calculation logic, a couple of timing bugs in the workout sequencing. I asked the AI for help. I got answers that looked plausible, made the changes, and introduced new bugs. Round and round. Eventually I put the project in a folder and told myself I'd come back to it later.

I nearly concluded from that experience that AI just couldn't build proper software yet. That turned out to be an accurate assessment at the time. It just wasn't the permanent truth.

What actually changed

When Claude Opus 4.7 came out, I went back and tried again. The bugs that had stopped me dead for weeks got sorted in an afternoon. Not because I had become a better developer. The model had just gotten better at holding the full context of a codebase in mind, reasoning about side effects, and catching the knock-on consequences of a change before suggesting it.

The bugs that stopped me cold a year earlier got sorted in an afternoon. Same problems. Different model.

Opus 4.8 widened that gap further. The key shift is not that AI writes perfect code on the first attempt, it doesn't. The shift is that the reasoning is now good enough that when something goes wrong, it can find why. It can hold the mental model of a complex system, spot the contradiction, and propose a fix that accounts for how the whole thing fits together. That is the specific capability that was missing before and is now present.

What I shipped since then

Since the models crossed that threshold, here is what I have actually built. No team. No funding. Mostly late nights and weekends at the start, now full-time.

Plus the client work: AI chatbots for cafes, booking systems for tradies, lead-capture tools for clinics, automations for small service businesses. Each one built solo, shipped in days to weeks rather than months.

The part I got wrong early on

I was too quick to decide AI couldn't build real software. That was a reasonable conclusion from the evidence at the time, but I made the mistake of treating it as a permanent truth about the technology rather than an observation about a specific point in time.

A lot of people made the same call. They tried GPT-3 or early Claude, got mediocre outputs on something complex, decided the technology was overhyped, and moved on. That conclusion was fair in 2023 and early 2024. It is not accurate anymore.

The specific thing that changed: the models can now handle genuine complexity. Not perfectly, not without iteration, but they can hold the context of a real system in mind, reason about tradeoffs, and debug their own mistakes when you point them back at the problem. That capability crosses a threshold where solo or small-team development of serious software becomes genuinely possible.

What this means if you are not a developer

If you run a small business, the same threshold applies to you, just shifted slightly. You do not need to build apps to benefit from it. The models that can now build complex software can also:

• Automate the repetitive admin that eats your afternoons
• Write first drafts of everything from emails to proposals to social posts
• Analyse your business data and tell you what is actually happening
• Run as a chatbot that handles customer enquiries 24 hours a day
• Connect your tools so information stops living in five different places

None of this requires you to become a developer. It requires either learning the tools yourself over a few weeks, or working with someone who builds these systems for a living and can have something running in your business in days.

The honest version of what AI still cannot do

It would be dishonest to write this post without saying what is still true: AI still gets things wrong. It needs context, clear direction, and iteration. The skill of working with AI effectively is real and takes time to develop. I still get plenty wrong myself and I do this all day every day.

The point is not that AI does the work perfectly and you just watch. The point is that the leverage ratio has shifted. The things that used to take a team of five people and six months now take one person and a few weeks. That is a genuine, meaningful change, not a promise of effortless output.

If you tried building something with AI 18 months ago and walked away disappointed, it might be worth another look. The version you gave up on is not the version that exists now.