I've spent most of my career trying to become a developer, but it never quite took. When AI tools and vibe coding reached what seemed like a practical stage, I figured this might finally be my chance. The idea was simple: build multiple SaaS side projects that could run quietly in the background while I kept my day job.
I built three of them. The market ate all three. The lessons were valuable, though it took me a while to see them that way.
The first one lasted about a week.
Right after ChatGPT became widely known, it had no memory. Every conversation started from scratch, which annoyed me enough that I built a wrapper to fix the problem. A week or so after I shipped my solution, OpenAI added persistent memory to ChatGPT itself.
I didn't think much of it. One data point isn't a pattern.
For the second project, I built an AI layer on top of Google Analytics and Search Console. GA4 was notoriously bad at surfacing anything useful, so the idea was straightforward: connect your account, talk to your web data instead of digging through reports, ask a question, get an answer, and have it email you a summary on a schedule.
Several other people were already building exactly that. Google came at the problem from the other direction, adding the ability to chat with your data directly inside GA4 and Looker Studio.
Two for two.
The third project was one I tested for months, which is why it stung.
It was a competitive intelligence tool I built for my day job. You give it a list of competitors and it watches their websites, social channels, press releases, and job postings. Every Monday morning, it emails a summary of what they've been doing. It looks at who they hired, what they changed about their pricing, and anything else that might represent a meaningful signal.
And it doesn't just list activity. It knows enough about the context of my organization to make a recommendation about what we might want to do next.
I looked at the market before I started. I knew the established players, including Klue and Crayon, were charging $20,000 to $40,000 a year. I also found one or two smaller projects on Reddit whose sites did not look like anyone involved had spent much time thinking about marketing or positioning.
I figured I had an edge.
Getting a working version running took a weekend or two of tinkering. It now operates inside the home AI lab I built, using Linux, an RTX 3090 with 24GB of VRAM, and local models for much of the routine work. Claude or Perplexity gets pulled in when something requires stronger reasoning.
Then I used it for a couple of months.
The reports landed in my work inbox every Monday morning. I'd read them, forward relevant pieces to other teams, and act on some of the recommendations when it made sense. Because the machine was sitting at home and not behind a corporate wall, I could reply to the report email with feedback and it would adjust the next one.
That part still feels like a small magic trick.
The system eventually became part of a broader effort to use AI for strategic decisions at the credit union, rather than limiting it to meeting summaries, first drafts, and other tactical work.
The product worked. It was viable, feasible, and something people might actually want. I started building toward a launch - creating a landing page, developing content, figuring out where potential users were. Maybe it would become a product. Maybe I’d hand it to friends in similar roles at other credit unions.
But I didn’t check again to see who else was selling it.
When I finally looked again, the market had shifted completely. No longer just two or three small projects - dozens, maybe closing in on a hundred, at every price point from $15 a month to $800.
One of them offered a Monday morning briefing covering ten competitors for $299. It promised intelligence that was interpreted, not just tracked, with recommended actions. That was close enough to the sentence on my unfinished landing page that I read it twice.
The head start I thought I had vanished while I wasn’t looking.
I killed the launch. I still run the tool for myself. I learned a ridiculous amount while building it, and I’ll probably hand it to a few friends who can get value from it.
It just isn’t a product worth pursuing.
This has happened three times now. I build something useful on top of a model, and within weeks or months, either the platform absorbs the feature and ships it for free, or a wave of people with the same tools build the same idea in the same window.
Vibe coding didn’t make me a bad builder. It made building easier for everyone at roughly the same time - that’s the shift worth paying attention to.
When the cost of building drops to a weekend for anyone with curiosity and a laptop, "I built something cool" is no longer quite as cool.
So what holds up?
I don't have a clean answer, but an audience is the one I'd bet on. Anyone can clone a tool in an afternoon. Nobody can clone a few thousand people who already trust you. Building that kind of audience is slow, unglamorous work that doesn't feel like building at all, which is probably why it stays scarce.
The hybrid setup is the part I'm less sure about. Local models next to frontier models, my own hardware, my own data, and no per-token anxiety still feels like an advantage. I genuinely don't know whether it is durable or just a longer head start, and I'd rather keep testing that than assume I know the answer.
Speed matters too. Once you land on something genuinely different, the window to capture market share is short and getting shorter.
The part I keep circling back to is that most of the value was never in selling any of this. It was in not paying a vendor and in getting noticeably better at my job by building the thing myself.
Before I reach for a software subscription now, my first question is whether I can build what I need. My second question is whether I can build it in a way that I'll still be able to maintain six months later.
The second question kills more ideas than the first.
Not everything has to become a business to be worth building.
None of this is a case against vibe coding. It has a real place, and I'm not trying to talk anyone out of it. If anything, my limiting factor isn't the tooling. It's that I know just enough about code to be dangerous.
Different problem, different post.