A • Open source
Six projects, shipped in the open
claude-desktop-debian alone: 5,318 stars, 56 contributors, 300 merged PRs, around 3,500 downloads a day.
See the repos →Non-Convex Labs
I make opaque systems legible.
I take systems that ship as black boxes and map how they actually work. Then I publish what I find: teardown reports with the raw data attached, agent definitions that include their failure modes, and open source that fills the gaps vendors leave. Non-Convex Labs is the imprint. The work is below, and it shows all of its receipts.
Shipping in the open
6 OSS projects
5,318 stars
~3,500 downloads / day
Every claim cited
01 • Where to start
Three ways into the work.
Pick the artifact that fits how you read: shipped code, a teardown with the data attached, or the writing behind the method.
B • Latest report
Inside the Code Split: A Chunk Census of Claude Desktop 1.19367.0
What each of the 44 satellite chunks in Claude Desktop's new code-split build is scoped to, with the strings that prove it.
Read the report →C • Latest writing
Can You Quantify a Writing Voice?
I fed nine years of my writing into a 25-dimension analysis pipeline. It handed back a spec sheet for my voice. With tolerances.
Read it →02 • Open source
The receipts are public.
- GitHub stars
- 5,318
- Contributors
- 56
- Merged PRs
- 300
- Downloads / day
- ~3,500
Figures for claude-desktop-debian, the flagship of six projects. See all repos →
03 • From the blog
Notes from inside the systems.
Building a Knowledge Base That Speaks to Both Humans and AI
Most knowledge bases serve one audience. Ours serves two through the same data: a web UI for humans and an MCP server for Claude Code. Same wikilinks, same search, same schema. Two interfaces, one system.
Redesigning an AI Document Pipeline: GPT-4o Vision to Gemini
A working AI pipeline was silently failing on 16% of documents, fabricating identifiers for 13% of records, and spending full extraction costs every run even when nothing had changed. Here's what the redesign looked like.