A threshold alarm fires on a value that has already been crossed. This reads the trajectory between the moments a human is scheduled to look, using a large language model and public federal drug data that any hospital can already reach for nothing. That method is the patent. Everything else on this site exists to show it survives contact with a real electronic health record.
Scroll the picture sideways to see all of it.
Hospital data on the left, public federal reference data underneath, one AI MedAgent Reasoning pass in the middle, and a clinician who decides.
Everyone in this business has heard a promising demonstration before. Here are the five objections that come up every time, and what was actually measured against each one.
Yes. Two certified production APIs, authenticated and read this month, and the results are published rather than summarized.
Epic grants system/Condition.read in the token and then answers that read with HTTP 403 and a zero byte body. The reason came out of an error string on a different endpoint: Epic authorizes below the OAuth scope, at sub-resource level. Their documentation does not say so. It was found by measurement.
Sandbox data is worse, and that turned out to be useful.
In Oracle Health’s open sandbox one patient’s recorded weight moves from 87.543 kg to 49.895 kg in one second, and 63% of her weights are an exact whole number of pounds. Epic’s densest hour is the same four normal values filed seven times in eight minutes, height included. Both were served by a certified API with no warning attached.
No threshold check catches either one. 45 kg and 89 kg are both plausible weights and 37 °C is a normal temperature. Only something reading the trajectory notices, which is the whole argument.
Correct, and that is why the second output exists.
The engine states, on every single call, how many of the eight standard medication safety checks it was able to perform and names the ones it could not. On the Oracle record that number is two of eight. When its allergy screen returned 775 text matches, it printed none of them and said why: a screen returning that many alerts has stopped being a screen, and the number is evidence about the method rather than about the patient.
An engine that reports its own blind spots is checkable. One that does not is a brochure.
Also correct, and the site says so rather than papering over it.
The socket for a commercial drug knowledge base is built and deliberately empty. Calling it throws DRUGKB_NOT_CONNECTED, in public. All eight screening classes are named on the page, and the ones running on public data alone are marked apart from the ones sitting dark.
Public sources carry this a long way. The page states precisely where they stop.
It is CDS Hooks, a published HL7 standard, live at aimedagent.net/cds-services. Any electronic health record that speaks it can call it, and building a service requires nobody’s permission.
On the evidence page the final step of the sequence is not a replay. It is a live call made from your own browser to the deployed endpoint, and the card that comes back is the card a clinician would see in the order screen.
Below is the working demonstration, open to anyone. One critically ill patient across a compressed encounter, with AI MedAgent Reasoning running between the scheduled human checks. Every device at the bedside feeds an interface engine, and a gateway streams that data with identifiers removed.
AI MedAgent now reads Oracle's FHIR servers as a registered app and takes live device data from an OpenBedside gateway running in the cloud. The pump is simulated and the patients are synthetic; everything else is real, and you can check it. What is real, and how to check · All changes, dated
Everything in the panel above is the reasoning. This is the sliver of it that actually arrives, on the device of whoever is covering at the time. CDS Hooks caps a card summary at under 140 characters, so that is the budget, and every card prints what it spent.
The consult note is not thrown away. In a real deployment it sits behind the card’s links entry as a SMART launch, which the clinician taps when they want it. On this page, the panel above is that launch.
Cards accumulate for the whole encounter, deliberately. By the end of a shift this is the pile, and the pile is why most alerts get dismissed unread. Anything added to it has to earn the space.
Start the simulation. The cards and this JSON fill in together.
Across this encounter, AI MedAgent flagged the risk well before the scheduled human checks would have caught it. That early catch, reasoned from evidence, is the method. Now the question that matters: does this hypothesis have merit? I want your honest read.