We build MCU CoPilot, an AI-assisted workflow that turns a medical check-up clinic's raw data into reports ready for doctor review. Most of the work, though, was never the AI. It was reading how a check-up clinic actually runs, then building a layer where staff, doctors, and AI each do the part that is theirs. Every report still goes to an authorized clinical reviewer before validation or distribution. Here is what we found in the workflow, and what we built on top of it.
Why we built it
A single corporate or institutional check-up can cover hundreds or thousands of people. Each person moves through multiple stations: blood tests, imaging, ECG, vision, hearing, and more. Results land in different systems, files, or paper records. Someone then consolidates each person's data by hand, a doctor reviews every record, and reports go out one at a time. Depending on the size, getting all of it out the door can take days to weeks.
What struck me was where the time actually goes — and most of it is not medicine. Clinic staff spend hours consolidating results, assembling reports, and preparing distribution. Doctors must still read and sign every report, so routine assembly competes with the cases that genuinely need deeper judgment. MCU CoPilot prepares the data and drafts a structured conclusion so the team is not rebuilding every report by hand. The doctor still reviews each report and makes the decision; the difference is that they do not have to reconstruct the case from raw files before they can judge it. That is why preparation and review time are worth measuring in a bounded workflow assessment. The same assessment should examine whether doctors receive clearer evidence and can direct appropriate attention to cases that need it.
Not AI for its own sake
The easy version is to point a large model at the problem and call it solved. That makes a good demo and a bad product. The hard parts are buried in the workflow, and you only see them once you have done the work by hand.
The same blood test may appear as HB in one lab's file, Hemoglobin in another, and Hgb in a third. MCU CoPilot can propose source-to-field matches by meaning, but every suggestion is reviewed and confirmed by an authorized user. Confirmed institution mappings can then be reused for matching headers. Records also have to be joined across files by the correct identifier and in the correct scope. If a source is missing, it stays missing; the model does not fill it in.
Validation is a human approval boundary. Basic and diagnostic editing is locked for the validated state. If source data changes, the record returns to correction, review, and validation, and a new PDF must be checked before distribution. MCU CoPilot works from documented inputs: it does not independently interpret raw X-ray images, ECG waveforms, audiograms, or spirometry traces. Those conclusions remain the work of the responsible specialist or technician.
There is no single definition of "normal"
This is the part we could not have designed from a whiteboard. Check-up institutions use different field sets, units, reference conventions, report templates, and occupational requirements. A reference range is an input to review, not an autonomous clinical answer.
MCU CoPilot respects approved institution mappings, references, and report instructions instead of treating model output as policy. Those settings shape the draft, but they do not replace source verification or clinical review. An absent value, threshold, or rule is not something the model gets to invent. The institution's approved guidance is controlled and reviewed, not free text we guess at.
"Fit for work" is a judgment about a job, not a number
Corporate check-ups make this concrete. Fitness cannot be inferred from one measurement alone; it has to be considered against a known role and the institution's approved occupational process. The same finding can raise different questions for an office worker, a driver, or someone working at height.
We learned to ask clinics to describe that context as they would brief a clinical colleague, not as engineers writing a prompt. Reviewed instructions can help MCU CoPilot prepare the relevant evidence and a consistent draft. They do not turn local text into an autonomous fitness decision. The responsible physician resolves the evidence, the role requirements, and any ambiguity, then decides and signs.
Why we test the way we do
If a doctor is going to sign the report, "it looked right once" is not enough. In evaluation, we run the same cases repeatedly and test whether required content remains complete, stable, and anchored to source findings. That evaluation informs release decisions; it does not replace the doctor's review of each production report.
We turned that method into a public benchmark: one fixed workflow, twelve foundation models, and thirty synthetic-regenerated Indonesian MCU cases tested in staged cohorts of 6, 15, and 30. Ten of twelve models cleared the 15-case machine-side candidate gate; nine cleared the full 30-case gate. This public report presents machine-side evaluation results only. It includes no local clinical-expert verdict and does not establish clinical correctness or production readiness.
Loop engineering, before it had a name
People are starting to call this "loop engineering." The idea is plain: do not ship a prompt and walk away. Use authorized, governed test cases and provenanced synthetic-regenerated fixtures; generate; check the output against explicit gates; find which layer failed; make the smallest fix; and lock that lesson into a regression test. One good-looking example is never enough to change a production path.
What deployment keeps teaching us
Deployment keeps bringing the conversation back to the same questions: can a conclusion be traced to its source, can the doctor see why a case was flagged, and does the approved report stay under human control? Faster output matters, but trust depends on those answers.
The hard problems are often ordinary. A value may arrive as 299 from one source and 299,000 from another. A large batch may need a sanity check before any draft is worth reviewing. We resolve those issues inside the approved institution workflow. When a pattern is safe to generalize, we turn the abstract rule — not the customer's record — into a reviewed mapping or test. The product gets better by making the workflow more reliable, not by treating customer data as reusable material.
AI writes the draft. Doctors decide. Our job is to make the source easier to inspect and the draft easier to review, while keeping the final judgment and signature exactly where they belong.
Micromeet — governed healthcare AI infrastructure
MCU CoPilot prepares check-up records and report drafts for authorized clinical review. Doctors verify, edit, validate, and sign before distribution. Each institution system, including its HIS, EMR, LIS, and booking system, always remains the source of truth for the records or statuses it owns. An approved and verified writeback only writes an authorized result back into that authoritative system.