01 — The waiting room
The engagement started in a hospital in Lyon at two in the morning. Our lead was shadowing a triage nurse who had been on shift for eleven hours. She had a stack of tablet EMRs on one side, a queue of forty-three patients on the other, and a triage protocol that hadn't been updated in six years. Across 42 hospitals in the group, wait times had crossed a threshold that made the front page of Le Monde. The board wanted an AI. The clinicians wanted anything but another dashboard.
02 — Clinicians in the loop, not the audience
We spent the first month refusing to write code. Instead we built a labeling console and put two staff clinicians inside it for six weeks, annotating de-identified EMR notes with the reasoning behind their own triage decisions. That corpus — 84,000 labeled encounters — became the spine of a fine-tuned clinical LLM. Crucially, every prediction the model made in training was reviewed by the same clinicians who wrote the labels. When they disagreed, the model lost, and the disagreement got logged as a permanent test case.
03 — On-prem, FHIR-native, auditable
European data residency is not a checkbox. We deployed inference on-prem inside each hospital's own network, wrapped in a FHIR-native ingestion layer so the assistant read the same structured records the clinicians did. Every recommendation carried a citation trail back to the source note. Every session was auditable. Every hospital had a physical kill switch on the console.
04 — The blinded trial
Before rollout, three pilot hospitals ran a twelve-week blinded evaluation: the AI's triage recommendations against the clinicians'. The AI won on accuracy in seven of the twelve weeks and tied in three. More importantly, the clinicians said it made them faster without making them feel replaced. Nine months after contract signature, Vitalis is live in all 42 hospitals, SOC 2 Type II and HIPAA certified, with zero PHI incidents and a 38% cut in average ER wait.