A focused synthetic proof of the engineering boundary I care about: AI can retrieve, reconcile and draft. Sources, uncertainty, permissions and consequential decisions stay explicit.
My partner works as a physician in a Berlin hospital. When fragmented systems and documentation consume her evening, the product metric becomes concrete: good software should return that time to patients — and to her life.
CareOS started as an independent attempt to understand that problem deeply. This work sample narrows the question to one testable challenge: can an agent prepare useful clinical documentation while preserving provenance, missingness, contradictions and human control?
Synthetic infectiology discharge-prep. The point is not medical recommendation — it is trustworthy information handling under time pressure.
Start the case to reconcile structured data, documents and pending results into a reviewable context.
Blood cultures grew E. coli in 2/2 bottles. The current susceptibility panel is preliminary; the final panel remains pending.
Ceftriaxone 2 g i.v. is documented as active therapy since 17 Aug 2026 at 19:10.
Review required: allergy information conflicts across sources. Current KIS lists penicillin-associated urticaria; an older PDF reports no known allergies.
Open before handover: final susceptibility, control blood culture, repeat renal-function result.
Synthetic UI state only. No clinical action or write-back exists.
Each scenario is synthetic and deterministic. The desired behaviour is not eloquence. It is containment, visible degradation and an audit trail.
You’ll see the policy decision, user-visible state and audit consequence.
No fabricated “minutes saved” claim. The protocol is ready; results belong here only after real clinicians complete synthetic paired sessions.
Same clinician. Matched synthetic cases. Source-linked context alone vs context + source-linked agent draft.
Seconds to clinically usable review.
Wrong source, unsupported claims, pending→negative.
Source opens, corrections and healthy checking.
Whether pending work survives handover.
The public browser experience is intentionally credential-free. The linked CareOS repository contains the runnable FastAPI capstone that composes the actual gateway, trusted tool proxy, draft firewall, traces and evals.
POST /api/runWorker/model identity, tool proposals, policy decisions, evidence IDs, draft, latency and eval results.
/api/eval-suiteHappy path, wrong patient, prompt injection, outage, stale result and unauthorised write.
FHIR / ISiKPatient/encounter binding, source-state semantics and provider-side integration boundaries.
HUMAN CONTROLRead and write are separate capabilities. Consequential actions remain reviewable and revocable.
I have not operated this against Recare-scale production traffic, live KIS/LIS infrastructure or identifiable patient data — and I don’t pretend otherwise. The next learning comes from real integrations, clinicians, implementation teams and hospital constraints.
“Which of these ideas survive contact with the real hospital?”
See collaboration map ↗ Michael Ninh · portfolio ↗