Notifyn started as a marketing plan for a different product.
I had built a small chat app for school and wanted to release it properly. There are already apps for this, of course, but I wanted one of my own, and I wanted it to be the most secure one I could make. So the question became the usual one: once it exists, how does anyone hear about it?
My answer was email, and the moment I tried to think it through, the plan stopped being simple.
Sending a lot of email is easy. Sending a lot of email that arrives is not. Send the same message to everyone and your bounce rate climbs, the filters notice, and you end up in spam. So each message would have to be different, written for the person receiving it. That much was obviously possible with AI.
Then a better question occurred to me. If I could change what each message said, could I also change when it arrived?
Everyone has habits. Someone who always opens email in the early evening should get it in the early evening. Someone who never opens anything on a Monday should get it on Tuesday. The data to work that out already exists in how a person has responded to past messages. A message written for one person, sent at the moment that person is most likely to read it: I remember thinking that would be a wonderful thing, and I dropped the chat app to build it instead.
Then the channels started multiplying.
Some people do not open email at all, so the message would need to reach them on WhatsApp. If WhatsApp, then SMS as well. And then the idea that made the whole thing feel enormous: what if it called them?
Imagine you spent an evening looking at car listings on a classifieds site and did not buy anything. A few days later your phone rings, and a voice asks, politely, what put you off, and whether any of these other cars might suit you better. I thought that would be genuinely useful, and I focused everything on it.
I named it Notifyn, because at the end of the day that is all it did: it notified people. I registered the domains, built a landing page, and built the portal behind it.
I have one principle I apply to almost everything: minimum cost, maximum speed, maximum quality. It is not always achievable, but it is the direction I start from.
Applied to voice, it meant not renting someone else's. Paying per minute for a commercial voice API was something I could not afford, and more to the point, something I did not want to depend on. Open voice models had become good enough to run myself, so I bought a workstation for training and inference and kept it running more or less permanently. It is still sitting there.
For a while it all felt possible.
Then I sat down and asked who would actually use it, and the answer took the project apart.
Notifyn could not be used by just anyone, because it only worked if you knew a great deal about the person on the other end. Not just their habits, but who they were.
Turkish makes this obvious. You cannot open a conversation with a stranger the way you would with a friend; the register is completely different, and getting it wrong is not a small mistake, it is the whole impression. You need to know how to address someone before you can say anything useful to them at all. A system writing to millions of people would need to know that for every one of them, and would need general knowledge of how people talk on top of whatever it knew about a particular person.
The phone calls made it harder again. People speak differently depending on where they are from. In some places, speaking loudly is simply how people talk. A system listening for emotion has to understand that, or it will decide a perfectly calm person is angry and respond accordingly, and then it has made things worse while sounding very confident about it.
None of these problems is unsolvable. All of them together, done properly, was not something I could build alone, and certainly not at zero cost. So I put it on the shelf.
It is still there, and I want to be accurate about what that means.
I did not shut Notifyn down the way I have closed other things. I did not decide it was the wrong thing to build. I decided I was the wrong size to build it: one person, no budget, a problem that needs data and people and money in roughly equal amounts. Those are different conclusions, and I think it matters to keep them apart.
Shelving it was hard. It knocked me for a while, mostly because I had no idea what to do next, and an empty week is uncomfortable when you have spent years making sure you never have one.
What I took from it is a question I now try to ask much earlier: not can I build this, which is nearly always yes, but who is this for, and what would they have to give me before it works for them? With Notifyn, the answer to the second half was everything, and I only asked after the domains were registered and the workstation was humming.
I have since heard the same lesson put more bluntly, by people who have built far more than I have: the building is the easy part, and the things that do not work are rarely the ones a single person could have saved by working harder. Notifyn was my first real encounter with that. It will not be the last thing I put on a shelf, but it is the first one I still look at.