Nairobi, Kenya · 01°17′S 36°49′E
WilliamObino
I train models and build the systems around them: sign language recognition, imitation learning for robot arms, and GPU compute for Intel Arc from Mojo.
Now
November 2026- shipping
- Ishara Soma v1, a Kenyan Sign Language to text model, in public preview since September 2026.
- porting
- Intel Arc GPU compute from Mojo through Level Zero, now a package on the Modular community channel.
- studying
- BSc Applied Physics and Computer Science at Daystar University, class of November 2026.
Selected work
16 projects in totalIshara Soma v1
Kenyan Sign Language to text. A landmark encoder routes each clip to one of four T5-small domain experts, with a phrase-retrieval layer for the sign vocabulary that generative decoding handles badly.
85.7%top-1 on 140 held-out clips
Mojo on Intel GPUs
A direct C-ABI Mojo binding to Intel's Level Zero runtime: device memory, kernel launch and dispatch, so Mojo can target Arc cards.
5.4 µsmin kernel dispatch on Arc B580
ACT for a humanoid arm
An end-to-end Action Chunking Transformer for pick-and-place on the ROBOTIS FFW humanoid in MuJoCo: ResNet-18 vision encoder, transformer encoder and decoder, conditional VAE.
30 Hzchunked control in MuJoCo simulation
LaunchGround
A hub for Kenyan builders: ship a project, get it reviewed in public, climb a leaderboard. Roast Mode, Toast Mode and the High Ground ranking are live.
livelaunchground.co.ke
About
Short versionI work where a model meets the machine that has to run it. A sign language recogniser that has to answer in a second on a CPU with no GPU; a Mojo binding that reaches Intel's GPU runtime with a C call's overhead; an imitation policy that has to survive contact with a simulator that is kinder than hardware.
I've also built products in Kenya: an AI tutoring platform (AcetutAI, site still up), a villa rental platform (Emerald Skyline, app launching soon) and LaunchGround. They taught me more about shipping than any side project.
The rest of the time goes into the open source that makes my own hardware usable.
- Studying
- BSc Applied Physics and Computer Science, Daystar University, 2023 to 2026
- Poster
- Top 50, Deep Learning Indaba 2026, Lagos: Mojo on Intel Arc through Level Zero
- Focus
- Imitation learning, visuotactile perception, GPU kernels, accessibility
- Languages
- English, Swahili
- Based
- Nairobi, Kenya
Writing
Notes from the workDots instead of pixels
How this site renders a photograph as two tones on a fixed lattice, and what broke when I pointed it at a dark-skinned portrait.
Soma v1: 85.7% top-1 retrieval, 0.34 WER
What Soma v1 measures on 140 held-out clips, why a mixture of experts suits Kenyan Sign Language, and where phrase retrieval does the work that generative decoding cannot.
5.4 µs dispatch: Level Zero from Mojo
Porting GPU compute to Intel Arc is mostly not a kernel problem. It is a question of who owns memory, who owns the queue, and what the runtime charges per launch.