David Briggeler I build Systeme that run without me.

by handruns

Starting point

Notes from my projects should appear on this website regularly. Alongside a full working week I do not write them all myself, and a language model must not invent anything along the way.

What I built

Once a week a script writes a field note from a topic pool in which every statement carries evidence from my projects. A language model does the wording, and the run discards any draft with a number or a link that is not in the pool. A second script builds the overview, the RSS feed and the sitemap, and wrangler from Cloudflare uploads the site. I get the draft on WhatsApp; unless I object, it goes online at 10:00.

What came of it

Until October 2026 Hermes wrote a weekday briefing here from eight news sources. The run of 23 July 2026 started at 07:00:01 and finished at 07:01:13; it found 18 items and used 3 of them. About three in four of its sources also appeared in the LOGBUCH on medienautomatisierung.com the same day. Since October 2026 Hermes therefore writes field notes from my own projects. You can read what comes out of it under Notes, with a date and evidence for every claim.

by handruns

Starting point

Eventforum Bern has no IT department of its own. When the network plays up, someone in the hall notices first, and nobody knows how long it has been that way.

What I built

A monitor queries the network every 30 minutes. It counts the connected devices and measures the load on the radio bands and the signal strength. When a value goes above the configured threshold, it sends me an email. A second watchdog calls a website I look after, and its interface, every five minutes. It only reports after two failures in a row and only once per outage, and it reports over WhatsApp. When the site is back, an all-clear follows.

What came of it

On 5 July 2026 the monitor ran during an event with 32 connected devices. Five emails came in that window, all of them benign. On 3 July my own laptop had triggered an alarm. I adjusted the thresholds twice after that, most recently on 6 July. So far the monitor mainly reports false alarms. A monitor that never reports is set up wrongly, and one that reports too often is ignored after two weeks. I am still calibrating.

by handruns

Starting point

Recordings of talks and events should appear edited and subtitled. Done by hand, editing, slips of the tongue and subtitles depend on one single person doing it on the side.

What I built

ElevenLabs Scribe transcribes the recording with word level timestamps. I cut out suspect passages and have them transcribed a second time, which is how I find slips that go unnoticed in the first pass. ffmpeg cuts according to the edit list. The subtitles come from the same timestamps, as do the on screen animations. I approve the edit plan before every run.

What came of it

Every edit list contains a _padding_params block with the values used: mid sentence 100 milliseconds after the last word and 80 before the next, at sentence boundaries 200 and 130 to 150, at the end of the video 600 to 700. Anyone who opens the file can see where each cut was placed. I do not quote a saving in hours, because I never timed the work before. I will do that on the next video.

by handmeasured

Starting point

Ads were running for a YouTube channel. The view count went up and the report looked good. Whether the viewers stayed was not in there.

What I built

I pulled the figures through the YouTube Analytics interface and split the views by origin, paid advertising against organic views. For both groups I calculated the average watch time and the reactions per view.

What came of it

By far the largest part of the views came from paid advertising, and that group stayed under a minute on average. The organic group was many times smaller and stayed many times longer. Since then I judge advertising by watch time and by reactions per view. The full calculation with all the figures is in the notes.

by handruns

Starting point

The board of Eventforum Bern expects figures every quarter. They sit in the accounts, and until now the report was produced by copying extracts into a document.

What I built

I pull the entries through the interface of the accounting software Bexio, in pages of 2000 entries, and remove the duplicates. A script maps them to the accounts and builds the report from that. I then check every figure again in the Bexio web interface, because the interface returns wrong values for some accounts.

What came of it

The mapping rules are held in the project and can be read. Every figure that comes from the interface traces back to an account. Figures that are estimates, the occupancy rate for example, are marked as estimates in the report. How long the earlier manual work took was never recorded.