JosueDiazFlores
Computer science student and builder. Small models, real hardware, work that runs where people actually are.
I learned most of this by building the thing before I was qualified to.
Grew up in San José, two years at De Anza, then Santa Clara. Lately: vision models on low-power chips, distilling language models down to something that fits, and the documentation nobody else wanted to write.
I care about work that runs where people are: on a sidewalk, on a pair of glasses, on a device with no signal. Theta Tau, ACM, a board seat at AI Kitchen. Currently reading Dune.
Projects.
Closed-loop navigation fully onboard a low-power chip. Three checkpoint generations, 60 MB to 52.
Won the compute, distilled the model from an open GPT architecture, fine-tuned it for tool calling.
Two vision-language models on Meta glasses, run in tandem, to spot people in the wearer's path.
Gemma, Qwen, Llama on 1,467 USCIS appeals. Only one configuration beat the always-dismiss baseline.
Where I've worked.
Trained a sidewalk segmentation model across three generations; built the onboard vision-navigation loop it drives.
On-device inference SDK. Production Claude Code skill, Swift release tooling, architecture documentation.
Failure-pattern reports for leadership, department-wide procedures, onboarding guides, led five interns. First-level support and ticket queues in ServiceNow and Remedy.
From here and from there.
Mexican-American, and not as a footnote. It's why I work in two languages, why the immigration paper exists, and why I'd rather build a tool that helps someone cross a street than one that helps someone scroll.
