Seven Marks: how a keyword-matching prototype became a hospital's AI platform
The version-by-version story of ETHER, from a February prototype whose AI was keyword matching to four autonomous agents in daily use at a public hospital network.
On February 22, 2026, I pushed a prototype called PIGI. It let anyone report a broken piece of infrastructure by scanning a QR code on the wall. It had three roles, reporters, technicians and admins, and Postgres row-level security keeping each one in its lane.
It also had a feature called "AI categorization". It was keyword matching. The README said so. That line matters more to me than any feature in the product, because every honest version after it was built on top of that first admission.
Seven months later the same idea, rebuilt and renamed ETHER, runs every day inside a multi-site public university hospital network. This is how it got there, one Mark at a time.
The seven Marks
- IFeb 22 · PIGI. QR reporting with no login, role workspaces, row-level security. "AI" by keyword, disclosed.
- IIApr 19 · ETHER v1.0. Rebuilt from scratch as an AI help desk for the hospital's engineering team.
- IIIApr 26 · Agentic. Agentic AI, "ETHER Brain", plus a security pass. One week after v1.0.
- IVMay 17 · WhatsApp. Technicians create, update and close tickets from WhatsApp. The same day, four medium findings from an OWASP ZAP scan were fixed.
- VMay 24 · Multi-department. Housekeeping, occupational safety and patient safety joined, each with its own flow. Nine issues from a full audit closed.
- VIAug 15 · Proof of isolation. A test suite that signs in as every role and checks what each one can read and write. It runs every night.
- VIISep 16 · Autonomous agents. A morning dispatcher, a deadline watcher, a closure auditor and a weekly management report. A network view across every hospital.
Rebuild when the foundation lies
Mark II was not an upgrade. It was a new codebase. PIGI proved people would scan a QR code instead of calling someone, which was the only thing it needed to prove. Its internals were built for a demo, not for a hospital. Carrying them forward would have meant carrying their shortcuts forward too.
The prototype's job is to answer one question cheaply. Once it has answered, its code has done its work.
Go where people already are
The change I am proudest of did not come from a smarter model. It came from Mark IV, when technicians could run their tickets from WhatsApp. They were never going to open another app between two repairs. The software had to come to them.
Great technology disappears into the day of the person using it.
Security is a test, not a promise
A hospital network trusts its maintenance platform with who reported what, where, and when. So from Mark VI on, the permission rules are not something I believe are correct. They are something a test proves every night by impersonating each role. If a change leaks data, the suite fails before anyone else notices.
Agents that work, not agents that chat
By Mark VII the AI stopped being a chat window. Four agents run on a schedule: one assigns the morning's work, one watches deadlines, one checks that closed tickets were actually signed off, and one writes the weekly report for management. The interesting part is not that they use a language model. It is that each one does a job someone used to do by hand, on time, every time.
What seven months taught me
- Ship the embarrassing version. Mark I was small and honest. That is why there was a Mark II.
- Version numbers are a promise. Each Mark had to be usable on its own, not a step toward some future release.
- Build next to the people who use it. I work in the hospital's engineering department. The feedback loop was a hallway, not a survey.
ETHER is in production at a public hospital network in Colombia. The dates above come from the repository history. Explore the seven Marks on the home page.