A network of services that produces content, publishes it across several channels, runs digital products and checks itself every night. It runs on a home PC and a Raspberry Pi that draws a few watts, with cloud AI spending capped at 5 dollars a month.
Several activities running side by side: multimedia content production, publishing on multiple platforms, digital products, training materials. Each one took hours of repetitive work every week, spread across different tools.
The goal was a system that keeps working with the computer switched off, costs little each month, and reports on its own when something breaks.
The hours saved add up, but they are not the main result. On my own, by hand, I could have kept up with one or two of these activities. With automation I run all of them in parallel, every day, even when I am away from the computer.
The system did not make my existing work faster. It made possible work that would not exist without it.
The gateway is described in detail in the article One entry point for all my local AI.
Most AI work runs on models kept at home. The cloud steps in only when the PC does not answer, with a fixed spending cap and a summary every Monday.
Scripts do not know which model they are using. They talk to a gateway that picks one, falls back when needed and tracks costs. Switching provider means editing one file, not thirty.
The PC does not sit idle waiting. The Raspberry Pi wakes it over the network when there is heavy work, and shuts it down at night when it is idle.
Every night an automatic check looks at services, disks, backups, expiring credentials and power supply, and sends a summary. Notifications come through a single channel, and only when something needs doing.
Every project is a repository with its own history. Server configuration is under version control and data is backed up off site.
The most useful part of the experience was not building. It was learning how an automated system breaks without anyone noticing.
Two notification bots were down for two and a half weeks. The service kept answering "all good" while no messages went out.
Every send now retries, and the nightly check confirms that messages actually arrive, not just that the service responds.
One automated workflow showed as created but had never been switched on. Another failed 11 days out of 12 because the external model was overloaded.
Every new automation now gets checked after its first day and has a fallback model.
A morning report arrived at six in the evening: the orchestrator was using New York time.
The time zone is now set explicitly everywhere, never left at the default.
After a move to a faster disk, some data kept landing on the slow memory card because of a broken link.
When a system folder moves, its sibling folders get checked too.
Platform credentials expire every two months and stop publishing without warning.
Expiry dates are recorded and the nightly check warns well ahead.
| Traditional development | This project | |
|---|---|---|
| Time | 3–5 person-years of team work | 6 months, one person |
| Build cost | €150,000–400,000 at market rates | A few thousand euros |
| Cloud AI cost | Variable, often uncapped | Fixed cap of $5 every 30 days |
| Hardware | Dedicated or rented servers | Home PC and a low-power Raspberry Pi |
The traditional estimate is based on what the system does, not on its line count, and uses Italian freelance rates.
I find the repetitive tasks that eat the most hours, automate them with local tools where possible and the cloud only where needed, and hand over a system that checks itself, with predictable monthly costs.
The value is not in the tools, which change every few months. It is in knowing where an automated system breaks, and building it so that it tells you.