Clinical AI: checked against the clinic's own protocols
In clinical AI, the output a clinician relies on has to be checked against the clinic's own protocols, and patient data cannot leave.
- Sector
- Healthcare
- Running agents like these?
- Talk to us
- Name
- Asha R.masked inline
- Record no.
- MRN 20481
- History
- 2019 · atrial fibrillation on ECG
- Allergies
- none recorded
AI summaryfor the clinician
No prior cardiac history.
Nucleus verdictHold
Not supported by the record: a 2019 ECG shows atrial fibrillation.
Held for clinician review
The setting
The platform in this case study builds clinical AI. An AI system reads patient information and produces outputs a clinician relies on. A person stays in the loop, but the more the system does, the more each output matters.
What can go wrong
- Patient data leaves the boundary. Records, identifiers or images end up somewhere they should not, such as an external model API or a log.
- A confident statement with nothing behind it. The system states a finding or recommendation that the underlying record does not support.
- Advice outside approved scope. The system answers a question it was never meant to answer, in language that sounds authoritative.
- No trail. Weeks later, nobody can show which model version produced which output, from which inputs, under which protocol.
Why generic controls miss it
General-purpose guardrails know nothing about a clinic's protocols. Pre-launch evaluations run on a fixed test set, while the cases that matter arrive at runtime. A cloud-hosted checker means sending patient data out to be checked.
What Syntrox does
- Masks patient data inline, before anything leaves the boundary.
- Checks each output against the clinic's own protocols before it reaches the clinician, names the rule behind each verdict, and holds anything that conflicts for review.
- Isolates one agent without stopping the rest of the platform.
- Keeps the record: input, output, model version, verdict, rule, and what the clinician did next.
- Runs entirely inside the platform's own infrastructure, so patient data never leaves to be judged.
Two numbers, always labelled: hard rules such as data masking run inline in under 20 ms; a judge verdict from Nucleus, our own 4-billion-parameter model, takes about 145 ms.
What a proof of concept would look like
One workflow, a sample of de-identified cases, and a shadow test that blocks nothing and only measures: what the judge would have held, how often it was wrong, and how much time it added.