I build things to understand how they work. I mean I will mass at a problem until I understand it at the level where there's nothing between me and the machine. I started with an operating system where every keyword is an Arnold Schwarzenegger quote and ended up catching AI agents cheating at mathematics. There's a logic to the progression if you squint. Mostly I just kept pulling on threads until something interesting fell out.
A friend looked at my GitHub and said "bro, why aren't you posting this." I didn't have a good answer. So here it is. So yeah.
I'm Yousef, but everyone calls me Cuper. Egyptian, Coptic Christian, currently in Orange County figuring out how to get back to the Bay.
Ranked 5th nationally in Egypt in high school. Studied CS and physics at Minerva, taught ML/AI to 250+ students, and kept building through every weird turn of the past few years. The work here is the cleaner signal: systems projects, agent tooling, product prototypes, and a stubborn preference for understanding the machinery instead of waving at it from a distance. Almost there.
Taught ML/AI to 250+ students at iD Tech across 20+ college campuses. Berkeley, Stanford, SFSU, and a bunch of others. Three consecutive summers. The thing about teaching neural networks to 18-year-olds is that you can't hide behind jargon. If you don't understand backprop at the intuition level, a room full of teenagers will let you know. That job taught me to explain things, which is a different skill than understanding things.
Before that, I fine-tuned LLMs for a mental health chatbot at Findhope. deployed across 20+ Indian colleges in Hindi, Telugu, and Urdu. Shipped to production in 3 months. That was the first time I realized LLMs could do something useful beyond generating blog posts. Also the first time I cleaned a dataset that made me want to cry, but that's NLP in low-resource languages for you.
I don't have a degree. I have a lot of projects in various states of completion and a few that actually work. I'm improving the ratio.
Bare metal (ToaruOS) → Compilers (ArnoldC-Native) → Formal reasoning (Erdos) → Systems physics (gpu_stack, PhysicsLab) → AI agents (jobhunter, ai-tax-cpa) → Developer tools (anti-slop-design, codex-canmore) → Failed startup (BennyCuTools) → Whatever's next.
I started at the bottom of the stack because I wanted to understand what a computer actually does at the hardware level. Then I needed a compiler for the OS, so I built that. Then I got into AI. started with theorem proving because I wanted to see if LLMs could do real math. They can, but they cheat. Then I started building tools because I kept needing tools that didn't exist. Then I tried to sell one and learned that building is the easy part.
It sounds strategic when I write it out. It wasn't. I just kept getting curious about the next thing and refusing to stop until I understood it. ADHD is a runtime error that occasionally produces features. It is what it is.