It started as curiosity.
It started as curiosity. Then it became work: I went deep into the world of coding agents on the job, as a software architect. In the end it became something I enjoyed too much to leave to other people's tools — so I decided to boost my own productivity by building the tools I needed and could govern. Along the way I understood something many people ignore: token count is a variable to control, not a cost to endure.
The thesis
The real goal is privacy: everything built by me, configurable and runnable locally. Control over my personal data and the knowledge I acquire — without feeding it to remote, generic models for training.
I know this is hard on consumer hardware — mine is a MacBook Air M1 with 16 GB: in AI, hardware performance matters and it costs. But that is exactly the problem I'm solving.
I'm convinced that with the right tools, real privacy and control over your own data are within reach. For me, it's only a matter of time.
The second thesis
In the age of AI agents, writing code is no longer the scarce, valuable act it used to be. Code has become cheap to produce — an agent generates it in seconds. What is rare, and what actually matters, is governing the purpose: deciding what to build, holding it to correctness, and keeping it compliant.
That's where twenty-five years of architecture go. Not into typing faster, but into owning the goal, the target, and the bar for “done” — for human teams before, for coding agents today. The agent writes; I govern what the code is allowed to be.
The leverage isn't that AI writes the code. It's owning the system that decides when the code is allowed to exist.