What engineers actually think about AI-generated code
My dissertation research in plain language: the question, how I'm approaching it, and what I can say so far about how engineers judge code they didn't write.
Coming soonA reconciliation service that took 20 minutes now takes 16 seconds.
The service reconciles a Nautobot source of truth against the live network estate. At 20 minutes a run, it was slow enough that nobody ran it on demand — which meant drift was discovered late, by hand, and usually by the person it had already inconvenienced. The goal was not a faster report. It was making reconciliation cheap enough to run whenever someone wanted an answer.
- Found the real cost, which was not the network The runtime was dominated by repeated work inside the service rather than by device or API latency, so profiling came before any tuning.
- Fixed the Nautobot ID-handling logic Identifier handling was producing incorrect matches as well as redundant work. Correctness and speed turned out to be the same bug: the service was comparing things it should have recognised as already equal.
- Stopped re-reading what had not changed A companion pipeline now stores a hash per file and skips unchanged preloads entirely, which took it from about 3.5 minutes to seconds.
20 min → 16 s
roughly 75× faster, and correct where it had previously been wrong