I'm a Software Engineering student who got tired of using things I couldn't explain.
So I stopped. I went back to the math. Now I'm rebuilding modern deep learning from scratch not wrapping libraries, not following tutorials, actually rebuilding it, one residual update at a time.
The long game is a frontier AI lab. But somewhere along the way I got pulled into something else: computational biology. The idea that the same mathematical frameworks we use to model language might help us model how living systems compute, adapt, and fail...... That intersection feels like where I want to spend the next decade. So now I'm chasing two things in parallel: a mechanical understanding of how transformers actually work under the hood, and enough biology to eventually sit at that boundary seriously.
I'm not moving fast. I'm moving carefully.
And while I work towards that, I'm investing in the people around me. I believe the next generation of AI researchers won't all come from the places they've historically come from. Part of my job, right now as a student, is to make sure that's true.
Currently: reading papers, breaking things in JAX, mentoring peers, and learning to ask better questions.
| Area | Tools |
|---|---|
| ML / Deep Learning |      |
| Languages |   |
| Web & Backend |   |
| Data & Infra |      |
The Residual Stream — building transformers from first principles
The premise is simple: if you can't rebuild it, you don't really understand it.
I'm building modern deep learning from the ground up in JAX — tokenizers, attention mechanisms, transformers, training loops, and scaling-law experiments — paired with the original research papers and honest weekly build logs. Every component is transparent and testable. When something breaks (and it does), I document why.
The goal isn't a clean repo that impresses people at a glance. It's the kind of deep, mechanical understanding that makes you useful in a room where the hard problems live.
Inside the repo:
- Transformer and tokenizer implementations from scratch in JAX
- Paper notes alongside source research PDFs
- Weekly build logs — including what broke and why
- Mathematical foundations: linear algebra, calculus, probability
- Reproducible experiments and original ablations
Python 3.11 · JAX · NumPy · Jupyter · Git
I care about who gets to be in the room. Not just who's already there.
A lot of students — especially here in Uganda and across Africa — have the raw ability to contribute to frontier AI but lack the environment that makes that possible: structured learning, peer accountability, access to the right resources, and someone who's a few steps ahead willing to turn around. I'm trying to be that person, even while I'm still figuring things out myself.
What that looks like in practice:
- Peer teaching — breaking down ML concepts (backprop, attention, optimization) for fellow students through study sessions and written explainers
- Building in public — The Residual Stream isn't just a learning project; it's a reference others can follow, fork, and build on
- Open source mindset — every repo I ship is documented for the person who comes after me, not just for me
- Community organising — working toward creating structured AI/ML learning spaces for students who don't have access to formal ML coursework
elton = {
"reading": [
"Attention Is All You Need",
"Scaling Laws for Neural Language Models",
"The Annotated Transformer"
],
"building": "Transformer from scratch — The Residual Stream",
"exploring": "Computational biology × ML: where sequences meet biology",
"learning": ["Backprop mechanics", "Tokenization", "Training dynamics"],
"community": "Peer teaching, building in public, making ML accessible in Uganda & Africa",
"aim": "Frontier AI lab — Anthropic, DeepMind, OpenAI or equivalent",
"open_to": ["Collaborations", "Research discussions", "Open source", "Mentorship", "Study groups"]
}If you're working on something hard and interesting (deep learning research, AI systems, computational biology, open source) — reach out. Always up for a good conversation.