About

I'm Jason Gardner, and I don't fit in a box. I've mostly stopped trying. I'm an artist at heart, an engineer by blood, and a scientist by rigor, which in practice means I design things, build them, wire them up, write all the code, and then do the statistics to find out whether any of it actually worked. The engineer solves the problem, and the scientist makes sure the answer is real.

Professionally, I'm a fixer. Hand me the problem everyone else gave up on, and I'll learn whatever it takes, build the tool that solves it, and show the rest of the team how to use it. I don't pick the answer first and bend the problem to fit it. I try things, fail, learn something, and keep going until something works, which is a polite way of saying I spend a lot of time falling down rabbit holes and climbing out covered in muck and trivia.

I spent about a decade as a network architect, designing secure communications over wire, wireless, radio, and satellite links. Then I moved into the lab, where my specialty was instrumental analysis: spectroscopy, signals, and the instruments themselves, which I often ended up designing, building, or repairing. Along the way I built an analytical lab from scratch and founded a personal-care products company.

These days a lot of my work happens alongside a collection of AI agents, some running on Claude, some on Codex, and an assortment of local models. We all write code, and I argue with all of them. The code I'm actually writing myself (not vibing) is mostly Python with a little JavaScript, plus C whenever there's a circuit board involved.

I'm reasonably handy at most of this, handy enough to be dangerous, and there's always someone who knows any given corner of it far better than I do. If that's you and I've got something wrong, tell me. I'd rather know.

What I'm working on

Better image classification and object recognition models, built with strategy instead of brute force. Serenity ML is my self-built, self-hosted workbench for active learning. A general-purpose local model makes a first bootstrap pass over the data, and after that every round works the same way. Serenity looks at where the current model is weak, ranks the unlabeled stockpile by how much each image is likely to help, and pushes the most useful ones to the top of my review queue. I label somewhere between 10 and 50 of them, fine-tune the model on the result, and the new model takes a short quiz on a held-out validation set. How it does on that quiz shows where it's strong and where it's still shaky, and that decides what gets pulled from the stockpile for the next round.

The point is to build better datasets faster, with less human labeling and smaller training sets made of images that matter to the model (which aren't always the ones that look important to a human). None of this is new; active learning goes back to the early 1990s. I just find it genuinely interesting, because it's finesse and strategy rather than throwing more data and compute at the problem, which seems to be where most open-weight model development is headed. Alongside it, Lineament is a smaller facial analysis toolkit that measures things like identity, facial geometry, and apparent age and expression across a batch of images.

Spectral analysis on data that never came near a sensor. The computational tricks I learned pulling peaks out of spectra (spectral methods, noise smoothing, and deconvolution) are good at finding a signal where plain statistics gives up, and I'm pointing them at other kinds of datasets to see what falls out. There will be essays on this.

A gimbal-mounted rig that detects, recognizes, and tracks objects, and works out how far away they are from a single lens in the infrared. It's my excuse to get better at mechanical engineering and mechatronics, and, let's face it, the thing will probably end up as a miniature C-RAM turret for Nerf darts. It should start showing up under Projects soon.

And a book on naturally occurring poisons and venoms: the plants, fungi, critters, and minerals that can quietly do you in, told through historical and anecdotal accounts of poisonings and the chemistry of why each one is toxic. Its companion site, poisongarden.org, is where I try out the entries and the pictures first. The pictures turned into a whole bloody project of their own, which I wrote up in Growing poisonous plants out of arithmetic.

What's here

Notes are the short ones: a derivation, a measurement, one idea worked through. Essays are the long ones, usually about how something got made or why I think what I think. Projects are things I build: 3D printing, CNC, electronics, and robotics. That shelf is empty for now, but the turret is coming.

Elsewhere

About the site

electrohead.net is a static Astro site running on Cloudflare Workers. I write everything in Obsidian, and a small publishing plugin moves it here. Maths is set with KaTeX at build time from the LaTeX in the source, so nothing in your browser is doing sums. Headings are set in Saira, body text in B612, and labels, tables, rulers, and code in B612 Mono, all served from Google Fonts. The look is borrowed from engineering drawings, which is why every page has a title block in the corner.