AI
Local models, agents, and making them useful
3I stopped renting my brain
A real AI assistant running on hardware you already own. No subscription, no cloud, no data leaving the house.
Build your own AI agent
An agent is a model with a chat window and a few safe tools it is allowed to use. The keys are the risk, not the model.
I made a music studio out of free software and AI
Generate original music on your own GPU, split it into stems, and master it. Open source the whole way, nothing leaving the machine.
Network
DNS, private overlays, and radio
3Kill every ad on your network
One small computer between your router and everything else. Works on the TV, the phone, the tablet, and your visitors.
A private network for all your machines
Reach every machine from anywhere with nothing exposed. Plus the routing gotcha that looks exactly like slow hardware.
Text with no internet and no carrier
Long range radio mesh you own end to end, so a message still goes through when the grid does not. Planned, not built yet.
Storage and media
Your own library, and backups that restore
2Own your movies again
Your own media server on hardware too old to sell. Nothing disappears from your library because a licensing deal expired.
Backups that actually work
Three copies, two kinds of storage, one off-site, and the restore test that proves it. A backup you have never restored is a wish.
Home lab
The house, the hosting, and the hardware
3A smart home that is not listening for someone else
Local control, local voice, no account required. The lights still work when the internet is down.
I put this site in my own house
This page is served off a small computer in a spare room, reachable without opening a single port. Ten dollars a year, all in.
Your machine is not dying. The thing watching it is cooking it.
Thermal throttling that looked like dying hardware was the metrics container holding a core around the clock.
Reliability
Proving a thing did what it said it did
9What I learned the hard way
The rules paid for by real mistakes, written down so a few of yours are free.
My assistant said it did something. It had not.
A model will narrate success for things it cannot do, and asking it nicely to stop is not a fix. The gate that is.
The worst alert is a green one
A job reported success every night while doing nothing at all. A check that can only ever say ok is not a check.
Prove the monitor is not the thing that is broken
Dozens of outage alerts for a service that was fine. A broken gauge does not read honest, it reads empty.
If your health check does real work, it will take down what it watches
The status endpoint that became the outage. The observer has to be cheap.
Never let an AI job spend money unattended
See it, project it and pause, check the environment not the config, and make the expensive path unreachable instead of discouraged.
Measure before you delete
Sixty files to clean up, three of them causing the problem, and a sweep that would have taken the credentials backup with it.
My research agent faked it for three weeks
A proxy quietly dropped its tools, so it could not research anything and invented the answers instead.
Read the log before you believe your own story
The remembered pattern was wrong, and two different causes were wearing the same costume.
Security
Keys, secrets, and closing what is open
3Every automation gets its own key
One key copied everywhere turns one small mistake into all of them. Scope each job so a leak is one account on one host.
Secrets leak from the boring places
Not a breach. A hardcoded password, a world-readable backup, and a token printed by a test stub that only half replaced the real thing.
Prove nothing is using it before you shut it off
Close the port, stop the service, keep the data. Verify first, make every step reversible.