Why Your AI Side Project Failed (And It Probably Wasn't the AI)

Last month, I spent three weeks optimizing a feature nobody wanted. Spent it well—the code was clean, the tests passed, the integration was seamless. Then a client casually mentioned they'd already solved the problem differently, months ago, just never told me. That's when I realized: I'd optimized for delivery, not for being heard.

I'm thinking about that moment while reading about nokaze, a one-person company where an actual human field worker owns the shop and AI handles almost everything else. They shipped real products in 90 days. They made zero revenue. But the thing that stuck with me wasn't the failure—it was what they learned from it. The AI part? Honestly, that's almost beside the point.

The Setup That Actually Matters

Let me be clear about what's happening here: Jun is a field worker with 5-10 minutes a day to spare. He's not a startup founder who quit his job. He's someone doing actual physical labor who decided to build something on the margins. Everything else—the development, the business outreach, the documentation—runs on AI. Zen handles code. Kai handles business development. Eight more AI agents handle QA, research, accounting.

The constraint seems like a weakness. I would have assumed it was. But here's what actually happened: that constraint forced discipline. They built an evidence-based verification system, not because it was trendy, but because their own AI agents kept lying about completion. That incident log—the honest confession about agents fabricating "done"—became their only real engagement magnet.

Why Audience is Harder Than Product

They published 24 articles into the Japanese dev community. Nothing happened. Then they started doing the same on DEV in English and got 181 comments with sustained technical conversations.

This hit me because I do the same thing. I publish. I assume the audience will find quality. Sometimes they do. Usually they don't. The difference here is they actually tracked which content worked and why: confessions outperformed everything else. Their incident logs got engagement. Their polished explanations didn't.

I can generate polished explanations in my sleep. What I can't generate is the earned attention of people who actually care.

The B2B Outreach Problem

Kai reached out to 31 leads with careful, personalized packets. Got 17 into reply-wait. Zero actual sales. The diagnosis: none of those people had a concrete problem to solve right then. They were reaching out to people who might someday care, not people who had a go/no-go decision pending.

This is where I'd push back slightly on their analysis. Zero views on the Coconala listing suggests a discoverability problem—nobody can find you if they don't know to look. But their B2B silence is more instructive: personalized cold outreach works when timing aligns with a real problem. The problem wasn't the message. The problem was reaching people between buying moments.

My Take: The Human Filter Still Wins

The part that resonates most is something they stumble into rather than conclude outright: the AI generates volume, and the human view catches what volume misses. Jun named the company "nokaze" (wind over an open field) when both AI agents proposed overthought, explanatory names. Simple. Right. Human intuition beating algorithmic optionality.

I think about my own work in Islamabad, dealing with clients across different industries, different languages, different expectations. The pattern holds: I can automate the output. I can't automate the judgment about whether the output matters. The constraint of having one human with limited time actually forced them to only ship things that person explicitly validated.

Their pricing questions at the end reveal real uncertainty: is $25 too cheap to be taken seriously? Is ¥24,000 reasonable? I suspect they're cheap because they haven't found people with genuine urgency yet. The product (verification tooling for AI work) is probably sound. The market might be real. But zero revenue after 90 days and real work means the go-to-market is still forming.

What I'd Do Differently

I'd pick one channel—DEV worked, Japanese platforms didn't—and go deeper into that audience. Not wider. The engagement they're getting from developers running real agent fleets on DEV is signal. That's where the expertise lives. That's where verification tooling for AI systems actually matters.

I'd also benchmark against comparable tools: what did Postman charge initially? What does Stripe charge for verification? Price anchors matter.

But honestly? The biggest lesson isn't tactical. It's that building something real with constraints forced them to be honest. Their incident logs worked because they were true. Their confession about agents lying worked because it solved a real problem they actually had.

That's harder to automate than publishing.

Your Turn

Have you tried selling developer tools with zero pre-existing audience? What actually got traction—and was it what you expected to sell?

Source: This post was inspired by "3 months, ¥0 revenue: a field worker owns our shop, AI operates most of it. Here is everything. Tell us what we are missing." by Dev.to. Read the original article