What we learn building and running software, AI and infrastructure. Read it before you learn it the hard way.
Practical notes from real projects: what worked, what broke, and why. No client names, no fluff.
Discover the lab.
7 documents found

10 NOV. 2026
We built something we had to switch off
An AI agent rarely fails by malice. It fails by retrying a delete, fanning out writes or looping on its own output. How to scope the tools, cap the cost and build kill switches that fire.

03 NOV. 2026
Sometimes the right build is no build
Before we price a build, we check whether it should exist at all. The four cases where the honest answer is not to build, and what to do instead.

20 OCT. 2026
Why every AI text sounds the same
Language models drift to the most typical sentence. Why temperature and "be creative" do not help, and what does: facts, examples, constraints and measurement.

20 OCT. 2026
Slop in, slop out: the data nobody labeled
What happens when models train on, or retrieve from, their own output. How the rare cases disappear first, and the labelling and dedup that keep a corpus clean.

19 OCT. 2026
The backup that restored perfectly and was wrong
A restore that finishes without errors proves the files were readable, not that the data is there. The checks that tell a real backup from a convincing one.

19 OCT. 2026
The human in the loop who only clicks approve
An approval step is a control only if the person can actually reject something. Why review turns into a rubber stamp, how to measure it, and how to design review that catches errors.
