The work, measured

Six figures. The first four are measurements I took, one per repository, folded under their titles. The last two are drawn from the repositories themselves, so they change when the code does.

Fig. 1 — storage-manager-swiftThe same scan, three waysOpen the figure

296 GB, 956k files, one machine. The gap between the lanes is the whole argument for building the tool.

.lane-fill { width: var(--final) !important; } .lane-head { left: var(--final) !important; opacity: 0.28; } .lane-time-live { display: none; } .lane-time::after { content: attr(data-final); } .lane-time { color: var(--ink); } .lane-mine .lane-time { color: var(--accent); font-weight: 600; }

du -skone lstat per file

0.0s

parallel read_dir + lstatthreads, same syscall

0.0s

storage-managergetattrlistbulk, batched

0.0s

10.3× faster than du, because batching directory attributes beats one syscall per file. Full method in the repo.

Fig. 2 — outloudKey release to text on screenOpen the figure

Voice dictation, measured end to end on an M4 Pro: from letting go of the hotkey to the words appearing. The transport decides the spread.

Aqua Voiceadvertised insert latency

0ms

outloud, clipboard fallbackunfocused app, paste

0ms

outloud, dictationaccessibility write, native field

0ms

outloud, edit-by-voiceselection rewritten in place

0ms

Fully local. Recognition is on-device, so the only latency left is the write into the field, and an accessibility write beats synthesized keystrokes. Full method in the repo.

Fig. 3 — wattsbarA menu bar app, before and afterOpen the figure

The same battery meter rewritten from Python and rumps into native Swift against IOKit. Memory at rest, measured in Activity Monitor.

Python + rumps41 MB bundle, py2app

0.0 MB

Swift + IOKit165 KB bundle

0.0 MB

5.8× less memory and a bundle 250× smaller, for the same number in the menu bar: live watts read from the SMC every second. Full method in the repo.

Fig. 4 — mp4trimOne clip, every Discord tierOpen the figure

A 51.5 s, 1440p60 AV1 OBS recording, 415 MB lossless. The same selection encoded to fit each upload limit; the budget is computed first and the result is measured after.

Nitro, 500 MB1440p60, source bitrate

0.0 MB

Lossless trimstream copy, Dolby Vision intact

0.0 MB

Nitro Basic, 50 MB1080p60, 7.0 Mbps

0.0 MB

Free, 10 MB540p30

0.0 MB

Budget is 93% of the limit minus audio, then encode, measure, and retry lower if it overshot. Under the limit every time, or it tells you to pick a shorter range. Full method in the repo.

Fig. 5 — Work

18 repositories, drawn to scale

Every tile is one repository, sized by its source bytes and split by language. 5.8 MB of code in total, measured from the GitHub API. 11 of them are private and appear as shape only: the language mix and volume are real, the names and contents are not published.

Fig. 6 — Languages

What the work is actually written in

5.77 MBacross 21 languages18 repositories