AI in production fails in three ways: a vendor goes dark, a job fails silently, data leaves your control. This is the record of the rulings I made against each one, on my own stack, with what they cost and what broke. Where something is still untested at scale, I say so.
Build automations with AI
Keep AI output deterministic: let the model decide, let code render
Model output you cannot reproduce is output you cannot sign: the same input comes back different, and nothing in it can be diffed. The rule here takes rendering away from the model, which returns JSON while code renders all a reader sees, and names the model swap that fixed French regressions.
Reliable Agent Systems
A framework upgrade blinded the guard, and only a deliberate mistake showed it
An upgrade can move the path a guard's lookup resolves, so the guard stops firing while the install still checks clean. This piece shows the provocation I run before any unattended chain and the wider audit that proved the guards live, and why an install that still passes is not proof.
Sovereign & local AI
Dictation without eavesdropping: on-device speech-to-text as a hard privacy line
Your dictation audio either stays on the device or it does not, and that line picked the stack. This piece follows the on-device choice once audio egress became the one disqualifying criterion, and what ruling out the cloud apps gave up without pricing.
