Blare
I build native tools and measure what they cost. Below: the same 296 GB scan, 956k files, three implementations.
10.3× faster than du, because batching directory attributes beats one syscall per file. Full method in the repo.
13 repositories, drawn to scale
Every tile is one repository, sized by its source bytes and split by language. 1.5 MB of code in total, measured from the GitHub API. 8 of them are private and appear as shape only: the language mix and volume are real, the names and contents are not published.
What the work is actually written in
- 01C#21.3%311 KB
- 02Go14.5%211 KB
- 03Nix13.2%193 KB
- 04Python10.1%148 KB
- 05JavaScript6.5%95 KB
- 06Svelte6.5%95 KB
- 07Shell5.9%87 KB
- 08PowerShell5.5%80 KB
- 09TypeScript4.6%67 KB
- 10Rust4.3%63 KB
- 11Just2.6%39 KB
- 12CSS2.0%29 KB
- 13Lua1.6%23 KB
- 14Swift1.0%15 KB
- 15HTML0.2%3 KB
- 16Batchfile0.2%3 KB
- 17Dockerfile0.0%694 B
- 18C0.0%185 B
If I can't measure it,
I don't claim it.

I write systems software and the tools around it. Most of what I build starts the same way: something on my own machine is slower or more fragile than it should be, and the fix turns out to live a layer below where people usually look.
The disk scanner was a syscall problem, not a UI problem. The video trimmer was a stream-copy problem, not an encoding problem. Both got faster by removing work, not by adding threads.
Right now I'm reading into machine learning properly: transformer internals, self-attention, how inference actually spends its time, and where agents break down. I use AI daily and want to understand the machine underneath it rather than treat it as a box that returns answers.