My daughter is three years old. She broke her toy guitar and wants a new one. Instead of another toy guitar, I bought her a real Ukulele. It’s small, lightweight, and perfect for a small girl. Besides, it sounds way better than the annoying toy guitar.
She loved it. But I’m the one who played with it the most. I don’t know how to play the ukulele. I was learning, and it was fun.
One Saturday morning, exactly at 7:15 I got an idea for an app. Have you seen karaoke apps? You sing karaoke while the lyrics on the screen scroll down. I wanted to build one like that for Ukulele players (for myself). They can see the lyrics and the Ukulele chords scrolling on the screen. They can sing along while playing the ukulele.
The best part: you can find Ukulele Chords for almost any song ever sung on the planet. It doesn’t have to be “Country Roads” or “Somewhere Over the Rainbow” all the time. You can type in any unpopular song; the app can still generate chords for it. I want to build it with an AI agent that finds the song on the internet (YouTube mainly) and uses that to infer chords. Thus, as long as the song you typed in the search bar is somewhere on the internet, you’ll find the chords for it.
I sat down in front of my computer, opened Google AI Studio, and told it what I want. My wife wanted me to join the family for breakfast, but I kept saying, ‘One moment, please’, borrowing more time. By 8:30, I had not only developed the app, but I also deployed it, bought a domain, linked it to my app, and shared the link on a Reddit thread.
Here’s what it looked like.
I am a professional programmer. I’ve been coding for the last fifteen years of my life. Trust me, if I had to develop this all by myself, I’d have taken a good six months. But I pulled it in about an hour, even without writing a single line of code.
Honestly, AI is amazing.
Later that day, I checked my Reddit post. I got a fantastic response. One redditor even said, ‘Sounds like the uke app I have always wanted!‘ I slept well that day, feeling that I finally built something the world wanted.
The next morning, I checked the stats on Google Analytics. I’ve got around 200 people who tried my app. That’s great! I then checked the billing tab in my Google Cloud Console. That’s where I felt that what I built was a flop.
I’ve spent $52 running the app just for a day. The AI agent is responsible for most of it. Every time a user visits the site, they might use this flashy feature to search for songs that they thought they would never find. And that shot the cost sky-high.
You could argue that’s normal for an app like this. I agree. But there was more than one reason to kill the app.
A business that fails at a unit level doesn’t make money
Before this, I’ve run two blogs myself. I’ve monetized them with display ads and affiliate links. From that experience, I can say, in general, you can earn about $10-$20 from every 1000 visits to your site. If you’ve got something to sell, you could aim for $30-$50. That means my 200 visitors would have earned at most $4. But at what cost? $52!
Unit economics fail, and that doesn’t sound good to me. Even if I buy a company’s stock, I look at how much it makes from every sale. That’s something I learned from reading the works of the greatest investors the world has ever seen, like Peter Lynch and Warren Buffett.
Some have said that the cost of building an app like this for 6 months would have been recouped. But mind that this $52 is the running cost of the app. Not the cost to develop it. I developed it for free. The Gemini free credits were sufficient for everything I wanted to do in this app. The running cost, however, would keep accumulating.
Losing $48 ($52-$4) a day would end up losing $17k a year unless I find other upsell opportunities.
Is the pain real?
What am I solving for?
The ability to generate Ukulele chords for literally any song on the internet.
Certainly, a lot of people would want this app. I figured it out from my own pain. I loved playing the ukulele, and as a beginner, I’d love to have something like this. AS A BEGINNER!
But what will happen once I master it? I probably infer the chords myself from my experience. I don’t need an app to do that. That means the market I initially envisioned was even smaller. It’s not any ukulele player; it’s beginners who struggle to infer chords and coordinate singing and playing together. Pros don’t need this.
Okay, but what about those beginners? Will they not use it enough? They would. But beginners usually tend to play a dozen songs that are universally considered the ‘beginner-friendly ukulele songs.’ If it’s for my daughter, it’d be something like “Baa Baa Black Sheep” or “The Wheels on the Bus Go Round and Round.” For grown-ups, perhaps the same old “Country Roads.” Popular songs are sufficient for beginners. Pros know how to infer chords for other songs.
But when I used AI to infer chords, I burned credits. Which is why the $52. It’s a costly whistle I hung on a genuinely helpful Christmas tree.
The problem is real, but the solution is overkill.
A bunch of statically generated pages would do the trick instead of a sophisticated AI agent. And that’s exactly what I have done now. If you visit ukuflow.com today, it’s a bunch of songs you can pick and play along to. Made for beginners, by a beginner.
Start small and iterate.
As someone who worked on many SCRUM projects, I should have known this better.
But I also think this is the smallest experiment I have run: failed with only a $52 bill, not $5k or $50k. I have also done the course correction of making the songs static instead of live AI generation. But I still want to talk about it because now that AI lets anyone build almost anything, more could go down a rabbit hole.
The best place to start is testing your critical assumption about your solution. For my Ukulele app, this would be that people would want to see lyrics and chords scrollling like they do in Karaoke apps. Even a single song page would have tested this assumption. That’s my minimumfeedback. I can decide whether to scale it up with more songs, add more features like AI generation, or drop the idea altogether
Final thoughts
At one point in my organization, I was the only person who coded in Python. Today, I also picked up Java, JavaScript, TypeScript, and a few others. But you know what? So does my mom (who has never written a single line of code).
When AI coding came in, I instantly lost my competitive advantage. Thankfully, I didn’t rely solely on coding. That’s my first lesson.
Ideas were always cheaper. Execution wasn’t. With AI, although it seems execution has become easier, planning still hasn’t. To succeed as a vibe-coder, or at least not to fail dramatically, we should be more agile. We should listen to the end users a lot more than we used to, then iterate and test fast. That’s the bigger lesson reinforced after seeing my vibe-coded app fail.
