# How Vecco Med Trains Medical AI.mp4
Imagine you are like in the back of an ambulance going 60 miles an hour.
The radio is just blaring with dispatch traffic. Uh a trauma patient is actively crashing and you are literally physically bracing yourself against the wall just so you don't fall over.
It's chaos.
Total chaos. And you have I mean exactly 4 seconds to calculate a critical drug dosage based on you know totally incomplete information.
now uh imagine trying to teach an artificial intelligence how to do that.
I mean, it feels almost impossible because uh when we normally talk about new AI products, we picture this like pristine white room perfection.
Yeah, exactly.
You know, a glowing node that just ingests perfectly clean data and spits out a flawless binary answer.
But out in the field, there is no clean data. Like none.
No, absolutely not.
So, welcome to the deep dive. Today, we are looking at a stack of internal white papers and architecture documents from a medical technology company. called Veco.
Yeah. And this is fascinating stuff.
It really is. Our mission today is to explore how they are, you know, completely flipping the script on AI in healthcare
because their goal isn't to build some uh autonomous robot doctor to replace human judgment. They are doing something much more grounded. They are compiling human experience starting right in the training room
which honestly completely upends the traditional Silicon Valley playbook.
Most tech companies, they want to rush their product directly to the final end user. you know, just to gather data as fast as possible. But Vecramed is like deliberately holding back.
Yeah. And I think whether you are managing a team or trying to understand how technology scales human learning or Yeah. I mean, you're just fascinated by how people perform under extreme pressure, this deep dive is going to give you an entirely new mental model for how expertise is actually built.
Definitely.
Because what really stands out in these source documents is their core emotional brand promise.
Oh, I love this part.
Yeah. It's not about tech for the of tech. It's about a concept they call earned competence. The whole point of their architecture is simply to help a human medic confidently say, "When people are in danger, I know what to do."
That phrase, uh, earned competence, it really anchors everything that they're building. It recognizes that expertise isn't something you can just, you know, download into a tablet or beam directly into a medic's brain.
Right. Like in the Matrix.
Exactly. Expertise requires friction. It requires repetition. But to understand why previous medical technology has failed so spectacularly in these highstakes environments, we really have to look at the actual physics of the job.
Yeah. And the sources are very explicit about this. Emergency medical work happens in motion.
Always in motion.
A medic isn't just staring at a patient in a vacuum. They are triaging physical symptoms, listening to a radio dispatcher, trying to mentally draft a hospital handoff report, and uh anticipating the patient's next physiological drop. All at the exact same time.
I mean, the cognitive load is staggering. At previous software solutions, they often force these practitioners to stop doing their actual life-saving work just to service the technology,
which is wild.
It is. The software essentially demanded, "Hold on, stop treating the patient for a second and click this drop-own menu so the AI can track your progress."
Right? Or on the other extreme, you have the tech industry pitching this utopian idea of an autonomous medic, a system that just magically eliminates It's human error altogether,
which is a fantasy.
Totally. And Vecommed throws both of those approaches out the window. Like they don't use hype terms like revolutionary AI anywhere in these papers.
Instead, they focus their entire architecture on this foundational unit they call the encounter.
Yeah. The encounter. They defined it as a changing state where people observe, decide, act, and adapt over time.
Okay. So, a process,
right? It is a dynamic episode of work. It is not just a final score. or a static data point.
I want to translate that corporate phrase though, dynamic episode of work into gritty reality.
Cuz that basically means like a medic slipping on a wet floor while trying to find a vein for an IV line, making a split-second decision based on an oxygen monitor that might actually be malfunctioning.
Precisely. The encounter is messy. It's a living breathing span of time. It's filled with, you know, hesitation, auditory interference, overlapping commands.
Okay. So, let's unpack this.
If a tech company wants to build an AI that actually understands all that mess, logic dictates they'd want to get their sensors into the ambulance immediately, right?
That's the usual playbook. Yeah,
you'd want to put cameras in the ER, gather that realworld chaotic data and train the model. But, if I'm reading this right, Vetomed explicitly refuses to start in the actual ambulance.
So, wait, if they can't go into the field because of safety and regulations, aren't they essentially locked out of the data they need? to build the AI?
Well, not quite. They aren't locked out. They just shifted their target to the one place where failure is actually encouraged.
Oh, interesting.
Yeah. If you look at the mechanics of machine learning, field deployment is a terrible place to build a foundational learning engine. Real emergencies are highly unpredictable. You just cannot control the variables to see what causes a specific outcome.
And there's the obvious roadblock, right? You cannot fail safely.
Like if your beta software crashes, gives a bad recommendation in a live trauma scenario, the stakes are irreversible.
The stakes are absolute. Plus, the regulatory approvals required to put experimental tech into live patient care. I mean, those take years.
So, if a company tries to start in the field, they are basically locking themselves into a slow, dangerous, and legally complex feedback loop. They are starving their own AI of the sheer volume of data it needs to actually learn.
So, they move to the simulation lab.
Yes. They call the training room an environment with with learning density.
Learning density. I like that.
It makes sense, right?
In a simulation, encounters happen back to back. You can run 50 complex trauma scenarios in the time it takes an operational field system to see just one comparable real world event.
And you have complete control over the chaos.
Total control.
Like an instructor can deliberately manipulate the scenario. They can pipe in the sound of a blaring siren. They can swap out an experienced team member for a total rookie. Or they can, you know, suddenly drop the dummy patients simulated blood pressure just to see how the medic reacts to the surprise.
Right? Error can be studied safely. The technologies real job in the training room is to help the instructional team see, understand, and improve the encounter using hard evidence
instead of just guessing.
Exactly. Rather than an instructor relying on memory reconstruction trying to remember if a medic hesitated for, say, 5 seconds or 15 seconds before pushing a drug, the system records the exact reality of the timeline. So, it's kind of like watching game film in sports rather than just looking at the final scoreboard.
That is a perfect analogy. Yes.
Okay. But wait, I want to pause on this idea of training an AI on learners though.
Because if I am building a system to eventually support live medical care, training it on people making mistakes feels, I don't know, totally backwards.
It sounds counterintuitive,
right? Wouldn't you want to feed the AI videos of the top trauma surgeons and the most elite flight medics in the world? Don't you want expert D? It seems intuitive to want expert data for sure, but the source documents introduce a massive paradigm shift here called learning from learners.
Learning from learners.
Yeah. While experts show you what perfect performance looks like, expert performance is highly intuitive. The internal steps they take, the micro decisions they make, those are completely invisible to an observer and therefore invisible to an AI.
Oh wow. Because an expert just walks into the room, glances at the ation instantly knows what to do without even realizing how they arrived at the conclusion.
Exactly. They just do it. Learners, on the other hand, reveal the transitions.
The transitions,
right? They show the exact moment they get overloaded by information. They'd reveal what specific monitor they look at first when things go wrong, how they try to recover from a mistake, and which specific interventions from an instructor actually pull them back on track.
That is fascinating.
It's a gold mine. That transition space is the gold mine. for a learning system. You cannot just train an AI on the final perfect end state. You have to train it on the messy path to get there.
Let's visualize this mechanism just to make it concrete.
Let's say there is a simulated pediatric asthma attack in a training lab. The dummy patient is crashing. A junior medic freezes up. They hesitate on whether to intubate. How is the Veco system actually pulling data from that hesitation?
Well, the system is capturing the physical reality of the room. It is pulling audio from ambient microphones to hear the silence of the hesitation or maybe the stuttering commands.
it is syncing with the dumy's vital signs to map the simulated patients deterioration directly against the medic's timeline of inaction. It captures the exact second the instructor has to step in and verbally prompt the medic.
But capturing that hesitation doesn't magically create a better medic, right?
No. No, it doesn't.
The documents outline a very deliberate operating cycle to turn that raw, messy transition data into actual human capability. They call it their operating cycle. Capture, learn, improve and act.
Let's look at how this cycle actually functions. First is capture which we just discussed with the asthma simulation, right?
But it is selective observation, not just indiscriminate massive recording. Then comes a really vital step, human review.
Human review. So a person is still in the loop
always. The system might generate an automated transcript of the scenario and flag an error, but AI transcripts could be highly confident and completely wrong.
Oh, absolutely. It might transcribe a medical term incorrectly or miss the context that the medic was actually doing a manual pulse check while looking away from the camera.
Which is exactly why Veco insists that uncertainty must remain visible in the software. The human instructor reviews the AI's breakdown of the encounter.
The instructor has the authority to inspect, correct, or outright reject the machine's output. Only then does the cycle move to the next phase, which is learned. So, the system compiles these reviewed events to find recurring failure modes. Maybe it notices that, I don't know, 80% of junior medics hesitate at the exact same moment in this specific asthma scenario when the oxygen alarm sounds.
Yes. And once it learns that pattern, we hit improve and act.
Okay. The final steps,
right? The instructor can change their debriefing strategy based on hard data. Instead of just saying, "Hey, you messed up the airway management," they can say, "The data shows you lost situational awareness right when the heart rate monitor started beeping faster. Let's run it again. And I want you to verbalize your focus when that alarm hits.
If the learning never changes the next encounter, the loop is incomplete. This cycle is the practical translation of Veco Med's strategic mantra. Every encounter becomes an asset.
Okay, stop right there because here's where it gets really interesting, but also a bit concerning. I have to throw a massive red flag on the word asset.
because to you and me and you know to most of the public, when a tech company calls human behavior an asset. That usually translates to we own your personal data. We are building a massive data lake and we are going to sell it to the highest bidder.
Right? It is a completely justified concern given the history of the tech industry. But the sources draw a hard invariant line on their data governance.
Yes. When Veco uses the word asset, they mean retained learning value. They do not mean data ownership. A raw audio file of a medic panicking in a simulation is not an ass. asset. A giant pile of undifferiated simulator data is not an asset.
So they aren't just hoarding the raw recordings of people's mistakes.
Not at all. The asset only emerges when lawful permitted observations are converted into trustworthy generalized knowledge.
Okay, I see.
Vecom explicitly states they do not automatically have the right to reuse, aggregate or sell encounter data. Everything depends on explicit governance, consent and purpose. Human authority and privacy boundaries are non-negotiable.
That is a massive distinction.
It is. The asset is the generalized lesson learned, not the identity or the raw mistake of the individual medic.
That clarification is crucial. The structural knowledge is the asset.
Now, let's look at how this scales because if a single training clinic in Ohio runs this capture, learn improve loop, they are going to get really good locally,
But what if one in instructor in Ohio discovers a brilliant way to teach a recovery path, while a clinic in Texas is struggling with the exact same pediatric asthma scenario. Traditional learning models mean Texas just has to suffer through the same mistakes independently.
Yeah. Local experience does not compound on its own. The instructor in Texas might take 6 months to independently rediscover the exact same recovery path that the Ohio clinic found yesterday,
which is just a huge waste of time and potential.
Exactly. So, Veco solves this isolation with a mechanism they call backhole.
Backhole. Uh, it sounds industrial, almost like supply chain terminology.
It does, but it's really a governed transport mechanism. Backhole takes all these distributed local encounters from different training sites and normalizes the data.
What does normalizing the data look like in practice? Because I imagine an advanced clinic in Ohio might have a $100,000 dummy with Bluetooth. vitals. While you know a rural clinic in Texas might just have a basic plastic mannequin and a stopwatch,
right? The hardware varies wildly.
Normalizing means stripping out those local hardware idiosyncrasies and translating the core timeline of the intervention into a universal language.
it looks at the sequence observe, decide, act. It standardizes the metric of a hesitation or a recovery so that the data from the high-tech clinic and the low tech clinic can be read by the exact same system. Back hall turns isolated clinics into a shared memory network.
And once that normalized data is pulled together, it enters what the architecture documents call the product space.
this is the model layer. And the phrasing they use here is fascinating. They emphasize that the product space doesn't summarize the experience, it compiles it.
The distinction between summarizing and compiling is everything here.
Yeah. Summarizing is just compressing information. If you To summarize a movie, you are just reading the back of the DVD box. You get a neat little paragraph of what happened,
Compiling in a software engineering sense transforms raw material into functional testable architecture. It is having access to the director's cut, the outtakes, and the scripts raw code and building a new engine out of it.
Okay. To use a different analogy, think of a massive multiplayer video game.
Okay, I'm with you.
If you play a ridiculously hard video game totally alone, you only learn from your own mistake. You fall in a trap, you die, you learn,
But if millions of players are playing and the game code compiles all the telemetry data on exactly where everyone falls into that trap, the collective we learns the boss's attack patterns instantly.
That is perfect.
The game compiles the failure data so the next player survives.
That analogy holds up beautifully in this context. Through this compilation process, the product space creates empirical models.
What kind of models?
Well, one of the most critical models it builds is something called a Timing distribution.
Timing distribution.
Yeah. The system learns across thousands of encounters when a 5-second delay in treating a trauma patient is totally harmless versus when a 5-second delay pushes the patient past the point of no return.
And that leads directly into the concept that honestly blew my mind in these sources, the recovery margin.
it is profound. The AI isn't just grading medics on a simple pass fail basis. It is literally mapping the geometry of human failure.
The geometry of failure.
That is exactly what it is doing. Recovery margin measures how much viable option space remains before an error is completely uncorrectable.
So like a point of no return,
If a medic administers the wrong fluid during a trauma code, how many minutes do they have to realize their mistake, calculate the counter medication, and push it before the patient crashes permanently?
It maps the buffer zone.
It maps the buffer zone. So, Veimemed's competitive moat isn't just a big pile of medical data,
right? Because anyone can get data.
Exactly. A competitor could just go buy a massive foundation model. Veimemed's moat is the history of disciplined learning. It is a continuously refined empirical model of human failure and how human performance actually recovers under pressure.
Okay. So, they have built this massive compounding asset of learning safely inside the simulation lab. They map the recovery margin They share the data globally through backhaul. But where does this go next? Does Veco just stay in the simulation lab forever? Do they just build the world's best training tool and stop?
No, they don't. The architecture documents outline what they call the conditional horizon.
Conditional horizon.
Yeah. It is a very strict two-step growth model. Step one is training capability. Step two is field capability. Moving out of the lab and into live ambulances, flight operations, and real world trauma centers.
But they use the word conditional very deliberately. It is not a guaranteed road map.
Right? It only happens if step one proves useful, if the data governance holds up, and if the whole system is economically viable. They have to earn the right to move to step two.
Earned competence. Right back to the beginning.
Exactly. But think about the structural advantage they have if they do move to live environments. The documents state the field system should not begin with an empty notebook.
Right. Because if you launch a brand new AI startup tomorrow and drop your shiny new tech into an ambulance. The notebook is completely empty,
totally blank.
The AI has no idea what real operational chaos looks like. But Vecommed arrives in the field with a notebook filled with years of compiled experience about human hesitation, timing distributions, and recovery margins.
They will already know what auditory interference sounds like when two paramedics are yelling over a siren.
They will already know the most common visual dist. ractions that cause a medic to miss a dropping oxygen level. The system will be uniquely tuned to the friction of the environment before it ever touches a real patient.
But even armed with all that compiled data, they are absolutely militant about their safety doctrine.
The field system is not an autonomous robot doctor. The sources are definitive. Qualified clinicians remain fully responsible for diagnosis and handoffs. The human is the final authority.
Right? And the system is designed with a non Nonblocking fallback.
Nonblocking fallback. What does that mean?
This is a crucial engineering principle for field operations. If the technology fails, if the tablet dies, if the AI hallucinates, or if the cellular network goes down in a rural area, the medic's workflow must be able to continue uninterrupted.
Oh, that makes so much sense.
The software cannot become a bottleneck that prevents patient care.
It can never be a crutch. So, looking at these version five documents, The ultimate vision isn't to replace the human instructor or the human medic.
Not at all. Veco exists entirely to support human authority, not to silently transfer that authority to a machine.
The technology is designed to surround the medic, to make the inherent uncertainty of trauma care inspectable, and to preserve those critical recovery margins we talked about. The human being stays firmly at the center of the encounter.
To wrap up this deep dive, we have covered a massive shift in how technology interacts with expertise. We started with the sheer chaos of a real world medical encounter and saw why Veco retreated into the safe sandbox of the simulation lab.
Right. And we explored how they capture human transitions learning from learners and feed those messy realities into the compounding loops of back hall and the product space.
And we looked at how compiling that history of human failure creates a conditional future for actual field medicine where the tech arrives already knowing the geometry of failure.
It is a disciplined two-step approach that completely rejects the overnight AI revolution narrative.
It really does. So, anytime you see a company claiming their new AI is going to magically revolutionize a highstakes human job overnight, remember Vecommed. Remember the value of earning competence first in a place where it is safe to fail.
Because the most valuable data isn't pristine perfection. It is the messy transitionary process of how we get there.
Which leaves me with a final thought for you to mle over. We have talked exclusively about emergency medicine today.
But if Veco successfully builds a machine that can compile human failure, map recovery margins, and return that as improved human capability, well, what other highstakes chaotic professions could use a prod wasting the valuable lessons hidden inside our daily mistakes?
Aviation, nuclear power, crisis management, the list is endless.
It really is something to think about. Thanks for joining us on on a steep dive.