Writing
Projects I’ve built with AI, and essays about what holds up.
The 52 that never was
The decline that rebuilt my cardio program never happened.
The switch I hope never fires
A proof-of-life system where both directions of failure, false alarm and silence, are catastrophic.
Uptime was the wrong question
If every check is green while five weeks of notes sit on one machine, what exactly is the dashboard measuring?
Mailroom writes back
Mailroom drafts replies in Gmail and learns from the versions I send.
Second opinion, same blind spot
Twelve real findings, twelve same-day fixes, and a regression neither frontier model caught.
The scout that pitched this post
It reads my week, ranks the stories, and interviews me before it drafts.
Rosie joins the Johnstons
Rosie reads Home Assistant sensors and controls devices when I ask in Discord.
Healthy in every snapshot, sick for days
My homelab ran on zero memory headroom for four days, and every live check said healthy.
Four servers, one weekend, and an API that no longer existed
The to-do list was supposed to be the easy one.
Buy nice or buy twice
A GPU hold that survived three reviews, and the 72 hours that broke it.
Once a quarter, I lie to my monitoring
I attacked my own safety net with seven fake faults, and two vanished without a trace.
Six seconds of microphone memory
Fixing my voice assistant's false wakes needed the audio from before the wake word, and nothing was recording it.
A prompt constraint is not a gate
The agent I told to use only the browser answered correctly without ever opening one.
The router that solved itself out of a job
My local model passed its final exam for a job that no longer existed.
I hadn't actually closed the door
I wrote up a security fix in the past tense. My own pre-publish check found the hole still open.
The upgrade my own referee refused
Eight hours of training, a loss curve that looked perfect, and the answer was still no.
Patchday
Transitioning from auto-updates to containers that now ask first, once a week.
The audit my AI coach never ran
Every session went to plan while the plan itself was failing me.
The package tracker that reads my email
My email already knew where every package was; the apps just couldn't hear it.
Eight pull requests to my manager's repo
Shipping code to the one reviewer who also reviews me.
The Bitter Lesson, applied to my own harness
I audited my agent fleet against a 2019 essay, and the essay won.
My red pen has a new student
Every Sunday at noon, my email agent rewrites itself.
Putting the coach's plan on my wrist
Garmin accepted every workout I sent it; the watch told a different story.
The guardrail Gmail wouldn't sell me
The Gmail permission that let my agent label mail also let it send.
The homelab grows a brain of its own
65 test cases, six models, one question: which self-hosted LLM will win the bake-off?
Letting the on-call engineer act
The agent can now fix what it finds; the guardrails approve the change, never the argument.
Giving my health coach a memory
Coach grew a memory, gathered its own data, and started to form opinions.
Rebuilding the family cloud on storage I trust
A one-way-door storage migration, and the offsite backup it caught broken for years.
A knowledge base my agents can read
Years of notes, one open protocol, and no app allowed to hold them hostage.
Rewriting the manual for its real reader
My documentation's heaviest reader pays by the token, so the manual went on a diet.
A brain transplant for the homelab
Moving the family server to new hardware with no dress rehearsal.
An employee handbook for my AI staff
Every AI session is a brilliant new hire with amnesia; this is the onboarding packet.
Building the health coach I never had
A daily AI coach that reads my training history and writes the day's session.
The terminal in my pocket
My AI work lives on a machine in the rack, keeps running when I leave my desk, and follows me onto my phone.
Teaching the house when not to act
The hard-earned first rule of my smart home: never overrule a human.
The on-call engineer I couldn't hire
When an alert fires at 2am, an AI investigates and I find out over coffee.
Writing the manual my homelab never had
I set out to write a manual for future me; my tools read it more than I do.
Onboarding an AI I didn't fully trust
How an AI earned an account on the server my family depends on.
Building a blog I'd actually keep up
The voice spec that teaches an AI to ghostwrite posts I'm still willing to sign.
Nothing in the archive matches that filter combination yet.