Building useful AI products and the teams that ship them
Turning machine learning research into production ready products for
some of the world's most advanced organizations. Lately: running
language models on live video at Matroid
VLMs in factoriesLetting operators ask their cameras questions
in plain language, and get answers with evidence.
Scaling the team
Our product and business roadmaps at Matroid are as ambitious as ever.
Get in touch if there's a
role you might be
interested in.
ReadingJames Joyce, Ulysses. Just finished
Ousterhout's A Philosophy of Software Design.
Selected work
matroid · case studies
PRODUCT LAUNCH
Matroid MVP
Problem
You're a small, pre-product team with an idea. What do you build
next?
What I did
As a founding engineer at Matroid, I shaped our initial feature
roadmap and was a core contributor, designing and coding the
model training & deployment functionality that continue to
define our product.
Outcome
0 -> a viable business with $10M+ ARR
MANUFACTURING
Visual inspection at line speed
Problem
Defects are rare, lines are fast, and the experts who know what
"bad" looks like aren't ML engineers.
What I did
Built a detector studio and interfaces for prompting models that
brought deployment cycles from months to minutes. Traditional
dataset creation & annotation are now a last resort.
Outcome
Defects are routinely detected within minutes of plugging in
cameras
GOVERNMENT
Mission-critical AI
Problem
Outdated tooling, distributed data, and high user turnover are
all challenges when working with the government.
What I did
Worked closely with stakeholders to build a product that meets
complex organizations where there at. Continued to iterate on
product abstractions until we had something that was truly no
training required.
Outcome
Fully operational in mission-critical operations, making
operator more efficient and effective
What I believe
on building, and on the work worth doing
i.
The hard stuff is still hard.
Software is nearly free to write, and yet most people aren't using
new apps products they weren't a year ago. Figuring out what to
build and how to create value is more important than ever. In my
work at Matroid, I've had a knack for seeing new research and
imagining the useful product that might be newly possible,
building interfaces for creating prompt-able without datasets and
enabling users to successfully ask more questions about their
visual media.
ii.
Writing is thinking, don't outsource it.
Ask AI for a first draft or writing feedback, and it'll put words
in your mouth. There's value in struggling with the blank page;
it's the process for figuring exactly what you think. Share what
you think: honest is better than polished.
iii.
Care about what you work on.
Some of the brightest technical minds of a generation have spent
their careers directly or indirectly working on serving more and
better advertisements. The results have been disastrous: broken
attention spans, mental health crises, a decline in literacy.
Consider what you're work is for. I'm motivated by technology that
makes the world safer, gets people home to their families faster,
and that doesn't keep them on their phones.
iv.
Experiment then productionize.
Figuring out what to build and executing at a production-level
quality bar are two different challenges, and I love both. Move
fast and experiment with new experiences with reference customers
who are excited about what might come next. But once a feature is
ready for production, top notch engineering is necessary to make
sure your product is one customers can rely on.
Experience
[2016] – now
Matroid
Director of Product Engineering
Helped take Matroid from pre-product to tens of millions in
ARR. I wrote large parts of the original MVP, and I've landed
more PRs, hired more engineers, and won more Hack Weeks than
anyone else here. Off the engineering track, I wrote the
original marketing copy, which set the company voice for years
and got copied across the industry. I've also taken the
company head shots.
[2013–16]
IXL Learning
Product design & management. Redesigned the analytics
suite teachers and students use.
named inventor · matroid, inc. · via google patents
20granted US patents
4pending applications
5patent families, 2025 → 2017
Newest first: language-model interfaces you just talk to, back to
the convolutional detectors you once trained by example. The same
path the field took.
20252 granted
Language interfaces for vision
Type what you're looking for: a language model turns it
into a filter over detections, or a search that highlights
objects in an image, with corrections feeding back into
retraining.
Compress hours of fixed-camera footage into a short video where
every detected instance plays at once, and answer plain-language
questions about a live stream.
A detection unit that records, recognizes and stores only what
matters in real time, and can spin up new detectors for
never-before-seen objects and search history for past
appearances.
Predators (violet) chase, prey flee, and the fastest prey live
long enough to reproduce. Now and then a baby is born ±0.25
faster or slower, founding a new strain. The fast strains take
over and the line climbs. A port of
a project I built years ago.
