En bref (français). Étudiant de 19 ans à l'Université d'Ottawa, né au Cameroun, résident permanent du Canada. Je construis des systèmes dont on peut vérifier chaque affirmation, et je cherche un stage ou une occasion de recherche en génie logiciel, en systèmes d'IA ou en génie biomédical.
I'm Stone, a 19-year-old student at the University of Ottawa. I entered university through the Interdisciplinary Arts program, and I am working on a transfer to the dual degree in Biomedical Mechanical Engineering and Computing Technology. Before that, I spent a year in Computer Systems Technology at Collège Boréal in Toronto. I was born in Cameroon and I'm a permanent resident of Canada.
I like building things whose claims can be checked: exact results where exactness is possible, pass/fail thresholds written down before the experiment, and a plain account of what did not work.
Looking for: an internship or a research opportunity in software engineering, AI systems or biomedical engineering.
Research: troncons
Serving gpt-oss-120b, a 63.4 GB model, on a laptop with 16 GB of RAM and an 8 GB RTX 4060. The weights stay compressed on disk and are decoded only when a computation needs them; the host working set peaks at 384 MB, and the GPU footprint is set by the card, not by the model.
- A synonymous codon for zero. MXFP4 has two codes for zero (+0 and −0), and no computation of the engines considered depends on the sign. I proved that folding −0 into +0 before entropy coding saves exactly P(V = 0) · H(S | V = 0) bit per weight, and checked it on all 180 expert tensors of gpt-oss-20b and gpt-oss-120b to within 1.7 × 10⁻⁴ bit.
- Lossless. In version 1.2 of the engine's file format, the fold removes 1.80 GB from gpt-oss-120b (3.26 % of its expert-weight bits); the engine's logits stay identical to the bit over 5,120 steps on both models.
- Measured honestly. On this laptop, v1.2 generates 2.890 tokens/s against 1.988 for v1.0 (×1.454; an independent judge measured ×1.442). The smaller file alone would allow at most +3.4 %; a crossed experiment traces the rest of the gain to the pattern of disk read requests on this machine, and the write-up says it is not a portable result.
- Written up in three technical volumes, in English and French. A short paper is in preparation; the code, the measurements and the volumes are public in troncons.
How I work: I directed this project with Claude Code. I set the goals, declared the success criteria before any measurement, and had every key result redone by independent reviewer agents before I believed it; the agents wrote most of the code and the reports.
| Project | What it is | Stack |
|---|---|---|
| recompose | Turning a scanned math book into typeset source that provably says the same thing: verification instruments tested by planting known defects, hypotheses written before results. | Python, NumPy |
| moderation-pipeline-prototype | Research prototype: where can an attacker exploit the seams between moderation stages (selective classification, Dung argumentation, bridging), and can they be defended? Each defense is measured against a naive baseline. | Python, pytest |
| vigie | A native Windows app for watching coding agents think, live: their reasoning, messages, tool calls and images. Built with the agents it watches. | Rust, wgpu, WGSL |
| veilleuse | A full-screen flip clock and floating study timer, with a hand-made glass effect that Windows does not provide, written up as a short paper. | Rust, wgpu, WGSL |
Also: discoveries, the admission-control technique from the moderation prototype written up as a paper, a practitioner guide and an explainer, in English and French.
- Represented Ontario at the 11th Parlement francophone des jeunes in Yaoundé, Cameroon (July 2026).
- Member of the Youth Climate Ambassador program with Skills for Change in Toronto.
Languages: French, English. Tools I use most: Python, Rust, TypeScript, CUDA, LaTeX, Git.
Contact: arielstone883 [at] gmail [dot] com