A system that runs 24/7, unattended and under control
A software lab that forms hypotheses on public market data, tests them, discards those that fail and builds the tools it is missing. It runs on its own on a server with ten services since 13 September 2026. No update goes live without my go-ahead. It is research: it handles no real money.
Why I show it
It is not an SME project. I show it because it proves something a company cares about: I can build systems that work on their own for weeks, with nobody watching the screen, and stay under control.
What I did
I set fixed rules: every idea becomes an experiment with criteria written in advance; it is frozen and put to the test; if something changes, a new version is created, never a hidden edit. AI agents wrote most of the code; I decided what to build, in which order, and approved every release. A second AI model acts as an independent reviewer.
After
- Ten services running unattended on a server since 13 September 2026, with Telegram alerts and a control console.
- 55 updates released with tests and automatic roll-back.
- 3,830 automated tests passed; an experiment memory of 174 tables.
What it can do for your company
The same method works wherever a system must run on its own and keep a record of everything: quality control, maintenance, recurring checks, automated reports.
Technical details, for those who want them
- 768 code revisions in 18 days; about 120,000 lines of Python in 41 packages; 3,830 tests passed at 30/9/2026.
- PostgreSQL with 174 tables; idempotent logging of experiments, results and decisions.
- 55 versioned releases with server-side tests, schema migration and roll-back.
- Governance: no release without the owner’s go-ahead, candidates frozen with a cryptographic fingerprint, sealed protocols, an automated maintainer with tracked fixes, a second AI model as reviewer.
The other systems I built.
Staff attendance with a WhatsApp message
Beforeattendance and messages in a WhatsApp group, hours counted by hand.
Afterthe employee writes “in”, the system records it, the administrator checks.
Google Ads and landing pages: −38% cost per lead
Beforecampaigns run by hand, occasional analysis, a generic landing page.
Aftera new landing page; every day the AI analyses and proposes, changes go live only after approval.
Do you have a similar problem?
Tell me in two lines: I will tell you whether the same approach applies to your company and what it would take to start.