MANAS
I build systems that work 24/7.
This is my actual desk.
Two screens, a keyboard, and the coffee. It's 3 a.m. in Waterloo, nobody's sitting here, and everything's still running.
I just wanted to know how it could talk back.
That was the whole hook. Not the industry, not the hype. Something on the other side of a text box was holding a conversation with me and I needed to know how.
Turns out nobody's home. It's math.
No little mind in there being clever. Matrices, probabilities, a very good guess about the next word. I sat with that for a while and went, okay. Okay.
I was good at math. This was not that math.
Top of the class in high school, so I figured I'd read the papers and get it. The math underneath this is on another level, and I'm honest enough to say going all the way down that hole isn't where I'm useful.
I learned exactly as much as I needed.
Transformers, embeddings, tokens, what a parameter actually changes, where machine learning ends and wishful thinking starts. Enough to build with it properly instead of poking it and hoping.
School had me writing a button.
Python, Java, write a small function, make a web page button work. It never connected to anything real. I could never make myself care about an exercise that helped nobody.
I build tools people actually use.
AI didn't make me lazy, it removed the ceiling. A barbershop owner runs his own site. A realtor gets her leads. A phone gets answered in Hindi at midnight. I'm not handing my work to AI, I'm using it to be better at mine.
Strong opinions, loosely held.
- Tools somebody actually uses on a Tuesday.
- Local models. My data staying on my machine.
- Handing a client the keys instead of an invoice.
- Showing up with a working demo, not a deck.
- Being trusted with the whole problem.
- Coffee, at an hour that worries people.
- Exercises that help nobody. I never cared about the button.
- $2,000 a month for something you could just own.
- AI strategy talk with no working software attached.
- Charging a client every time they want to change a price.
- Demos that only work in the demo.
- Pretending the AI did none of it. Or all of it.
