A pending result stays pending.
“Not back yet” never becomes “negative” or “normal.”
Recare work sampleAI / ML Engineer
Give clinicians time back without weakening trust.
A synthetic case about a practical engineering question: can AI bring scattered clinical information together, prepare useful work, and still make uncertainty, sources and human authority obvious?
Source-linkedFHIR-awareHuman reviewBounded AIFailure replay
Four rules
What the system refuses to blur.
A useful clinical assistant should make these states easier to see, not harder.
If a source is down, say so.
Unavailable information is never silently treated as absent.
Charted treatment is not AI advice.
CareOS can show what is documented without pretending the model prescribed it.
An AI draft stays a draft.
Consequential work stays reviewable until a human explicitly approves it.
Why this problem
The cost of bad clinical software shows up after the shift ends.
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 is my independent attempt to understand that boundary deeply. This case asks one narrow question: can we prepare useful work while preserving source links, pending work, contradictions and human control?
01 · Guided clinical walkthrough
One patient. Five sources. One reviewable next step.
This walkthrough does not measure time saved. It shows how CareOS handles information and authority. The separate study below measures speed against a baseline.
- 1Startassemble the context
- 2Reviewresolve a contradiction
- 3Approvereview the AI draft
- 4Tracesee what happened
- 5Safety checktest one failure
Start with scattered clinical information.
CareOS assembles the relevant context while keeping pending work and conflicting sources visible.
AI-prepared · not source truth
human review requiredDischarge-prep summary
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 only. Approval here does not write to any clinical system.
What happened behind the interface?Every meaningful read, policy decision and human action stays inspectable.
0 eventsStart the walkthrough to populate the trace.
02 · Safety checks
When something goes wrong, does CareOS fail safely?
Choose one scenario. We are checking whether unsafe or uncertain input is blocked, quarantined or shown as degraded — instead of being turned into a confident clinical answer.
Safety check
Pick one failure scenario.
The result will show three things: the system rule, what the clinician sees, and what gets recorded for audit.
A safe stop is a successful test. It proves containment — not time savings.Next: how we measure usefulness ↓
03 · Measurement
Does CareOS actually beat today’s fragmented workflow?
This is where time enters the story. The same participant completes the same clinical task twice with matched synthetic cases.
Baseline
Fragmented workflow
Simulated KIS, LIS, medication, lab and document views. The participant must reconstruct the picture.
VS
CareOS
Unified source-linked context
The same task, with relevant information assembled and original sources still inspectable.
Only if it is faster without becoming less safe.
We compare completion time and workflow friction. Wrong answers, missed pending work or collapsed source verification can invalidate a speed gain.
Task time
Timer starts when the case appears and stops when the participant finishes the requested handover.
Workflow friction
Systems opened, context switches and repeated source hunting.
Verification
Whether the user still checks important source information.
Safety
Wrong answers, missed open work and explicit stop conditions.
Evidence state todayStudy instrument ready · real participant evidence still pending
Open clinician study instrument ↗04 · Engineering
The AI does not own clinical authority.
The browser walkthrough is deliberately synthetic and credential-free. The CareOS repository contains the runnable implementation, tests and explicit limits behind the interaction.
1 · Clinical systems
KIS · LIS · FHIR · documents
Existing systems remain the authoritative sources.
↓
2 · CareOS context layer
What do we know — and where did it come from?
Patient identity · source · clinical time · pending/final state · contradictions
↓
3 · Safety & permission layer
What is this user or AI actually allowed to do?
Patient scope · tool permissions · read/write separation · budgets · audit
↓
4A
Clinician interface
Source-linked context and human review.
4B
Bounded AI assistant
May propose work only inside delegated permissions.
POST /api/runExecution path
Worker identity, proposed tools, policy decisions, evidence IDs, draft and eval result.
/api/eval-suiteReplayable safety cases
Wrong patient, prompt injection, outage, stale result and forbidden write.
FHIR / ISiKInteroperability boundary
Patient/encounter binding and source-state semantics are explicit rather than left to the model.
HUMAN CONTROLReview before consequence
Read and write are different capabilities; approval remains revocable and auditable.
For HR: you can stop here. For engineers: the implementation guide shows how to run the API and safety suite locally; the repository contains the deeper architecture and tests.
The honest boundary
I took the synthetic problem as far as I responsibly could from outside a hospital.
I have not operated this against Recare-scale production traffic, live KIS/LIS infrastructure or identifiable patient data. The next useful learning comes from real integrations, clinicians and implementation constraints.
“Which of these ideas survive contact with the real hospital?”
See Recare collaboration questions ↗Michael Ninh · portfolio ↗