emailmanager.ailive
Answers customer support mail
Drafting, classification, escalation rules
VIENNA –:–
last pass —
next run —
While I sleep, fourteen agents work my projects.
The schedule below is the one that actually runs, times taken from the config. They research, write, grade their own output and file it. Nothing goes out until I say so — that part is still me, and it is the part I am least willing to automate.
Scheduled, not measured. This board reads the timetable, not last night’s logs. When a run fails you read about it in What broke, not here.
Fourteen runs across eighty-five minutes. One dot per run; where two share a minute, they run side by side. The marker appears only while the night is actually under way.
doctrine-review
Rereads my own doctrine files and flags the rules that stopped matching reality
plan-critic
Argues with last week’s plan and says where it is now wrong
contrarian
Takes the opposite side of whatever I decided, on purpose
diagnostics
Checks the whole machine for silent failures — the watchdogs that died quietly
analytics
Pulls traffic and search data into the ledger
seo-blog
Researches, writes and files a full article draft
pr-cycle
Works the press cycle: who to pitch, what changed, what is stale
trend-briefing
Reads the feeds and writes the briefing I read at breakfast
linkedin
Drafts the LinkedIn post
community-reddit
Drafts the Reddit answer and checks it against the sub rules
community-x
Drafts the X post
output-grader
Grades last night’s output and says what was weak
idea-spark
One idea, argued through, filed as a card
sales-inbox
Sorts the inbox and drafts replies for me to approve
By morning there are drafts, a briefing and a grade on last night’s work. I read, cut, and press the one button they are not allowed to press.
The work
Five are products, five are the machinery that keeps them fed, and one is a gym log with no automation in it at all. All of it runs on hardware I own, and every claim on this site can be traced to a log on it. The pictures are taken by a script, on a phone-sized screen, and retaken whenever something changes.
seitenhebel.delive
Audits a website and rebuilds it
Crawl, scoring, revenue estimate, the whole report
swiftrun.ailive
Self-hosted platform for scheduled agents
Pipelines with connectors, a schedule and a report at the end
realquantum.ailive
My other log: quantum computing and AI
Research, drafting, publishing
tubewiser.ailive
Reads my YouTube backlog so I do not have to
Transcripts, briefing cards, scoring against my goals
The gym loglive
One HTML file, no server, no account
Nothing. This is the one I keep by hand
The blog factory
One pipeline behind five sites
Topic plan, research, draft, image, push
The cockpit
Where every draft waits for me
Tickets, deals, costs, the run log above
The video line
Script to rendered video
Voice, composition, render, cost gate
The home server
A ten-year-old ThinkPad under the desk
Every nightly run above happens here
The cost cockpit
What every model costs, and how much is left
Counting, projecting, and the push before the wall
The log
Four strands. Failures get the same room as the wins, because they are the part nobody else writes down.
-
The reconnect quietly swapped my channel id
Reconnecting X fixed its 280-character limit and silently gave the channel a new id. My config kept pointing at the old one.
-
The limit came from the docs, not the endpoint
My code said a post could be 25000 characters. The service in between accepted 280, and it had said so in a rejection nobody read.
-
Twenty-one cards, one hundred rejected
I sorted the heap of 121 open proposals my agents had made. 21 became work, 100 were rejected, and each rejection got a written reason.
-
The project switch was saved and invisible
I switched a card’s project 20 times. Every switch was saved, and the history showed each one as an empty note. The bug sat where the…
-
My cap switched off the dispatcher
A limit I wrote to keep one failure cheap stopped every card for 142 minutes, while three indicators stayed green.
-
My schedule test started a real run
I tested a cron schedule by calling the script the cron calls. It ran for real. What 14 crontab lines taught me about dry runs.
-
My 48-hour veto window ran for 72
A deadline set to 48 hours ran for 72, because the job that enforced it only looked once a night.
Who is doing this
I am Georg Singer. Physics degree, doctorate in computer science. Everything on this site I built myself, in my own time, on hardware I own.
Why this exists at all: I wanted to find out how far it actually goes — whether running my own products can be handed to software almost end to end, from the research and the writing to the bill at the end of the month. The only honest way to test that was on my own things rather than on somebody else’s company, so I am my own experiment. What I am really collecting is the list of what does not work, because that is the half nobody publishes.
Everything here is first-hand. If a number appears, it came out of a log on my machine, and if I could not measure something I write that instead of estimating it. That is the whole editorial policy.
Vienna, Austria9 systems running14 nightly agents0 of them may publish
How each of these gets built
The same three steps on every system above. Not a method I recommend — the order I actually work in, written down because most of what breaks happens between the steps rather than inside them.
first
Measure
Before anything gets built I follow the process as it really happens and write down what it costs me today. Every one of my expensive mistakes started with skipping this and estimating instead.
then
Build
Only the one or two steps that pay for themselves first. On hardware I own, with logs that prove the run happened rather than a dashboard that says it did.
after
Keep it running
Automation rots. Every month the watchdogs get tested, the rules get re-measured against what is true now, and the costs get read out loud. Most of What broke is this step failing.