About the imprint
Why "Non-Convex"?
In optimization, a convex problem is the easy kind. One clean minimum. One right answer. Gradient descent walks you straight there. Textbook stuff.
Real problems are not convex.
Real problems have local minima that look like solutions but aren't. Saddle points where progress stalls. Landscapes where the obvious path leads somewhere mediocre and the good answer is hiding in a direction nobody thought to look.
That's the work I do. The problems that don't have a clean slide deck answer. I named the company after the hard problems because those are the only ones worth solving. Non-Convex Labs is the imprint; the person doing the work is me.
01 • Method
The through-line
Everything I publish follows the same pattern: take something opaque, make it legible, ship the artifact with its receipts, and document the failure modes.
Claude Desktop ships as a minified blob, so I took a census of all 44 of its code-split chunks and published the raw JSON next to the report, so nobody has to trust the prose. My agent definitions ship with the ways they broke in real use written down. claude-desktop-debian exists because official Linux support doesn't: 56 contributors and 300 merged PRs later, it's how thousands of people run Claude Desktop every day, and in July 2026 Anthropic accepted it into the Claude for Open Source program.
Before software, I ran engineering and manufacturing programs: schedules, risk registers, root-cause analysis on hardware that couldn't be patched after it shipped. That's where the habit of receipts comes from. In that world, a claim without evidence is a slip waiting to be discovered.
02 • Standing rules
What I believe
Ship first, polish later. A working system that handles 80% of cases today beats a perfect architecture that's six months out. Get something into production fast, then iterate on what matters.
Be honest about what AI can't do. LLMs are powerful. They are also unreliable, expensive, and bad at math. I will tell you when a regex is the right answer. I will tell you when your problem doesn't need AI at all.
Pragmatism over purity. I don't care about architectural elegance for its own sake. I care about systems that work, that can be maintained, and that don't fall over at 2am. If the pragmatic choice is ugly, I'll write a comment explaining why and move on.
No black boxes. Legibility is the whole point. When I build something, you can see how it works: decisions documented, trade-offs explained, raw data downloadable. That standard applies to my own work first.
The person who scopes the work is the person who does the work. Non-Convex Labs is one person. What you see published here is what I actually ship, not a portfolio assembled by a team you'll never meet.