First ever Cherry meets Cornell Tech live coaching session. So for our listeners, what are we doing here? We’re actually on Roosevelt Island. And if you don’t know, but Roosevelt is actually famous for how he handled innovation.

So it’s actually very, very cool that Cornell Tech is here. We have three startups and we are doing something which is a live coaching session. This is all pre-seed. This is all early.

And all we do here is we chat about AI, data, the startups and how they fit. This is amazing. So let’s get going. Let’s first start with a quick round of introduction.

So my name is Oyebek and I am a Runway Postdoc Fellow. I was brought in here on fellowship to commercialize my PhD technology that I developed in grad school. So what’s Runway? So Runway essentially is an entrepreneurship fellowship at Cornell Tech that brings in PhDs and I think we accept five a year.

It’s a two-year paid fellowship to commercialize a technology that ideally you develop in graduate school. Pretty cool. I’m Shang. I am connected to Cornell through my PhD in computer systems, where I basically focus on building AI systems.

And I brought that expertise to build my current company called Zenith. We’ve known each other things like six years now, me and Lutz. We’re pretty good friends. I’m Vince Hartman, co-founder and CEO of Abstractive Health.

I’ve been in healthcare about 10 years. So I worked at Epic for about two, and then I did healthcare consulting for another six. And Abstractive Health, we’re very focused on solving physician burnout. So we take the data from a patient’s medical record.

And then we use it. So, yeah, it’s really cool. And then we summarize that contract. So we’re able to retrieve it from across the United States to create a real-time summary for physicians.

I want to give everybody a pitching moment, like three minutes pitch. Okay. So we are Neuralens and we quantify brain health using light. So we apply advanced methods of measuring how light interferes with tissue.

And in this case, we deal with measuring brain tissue specifically. We use it to quantify things like blood flow, blood volume, and intracranial pressure non-invasively in patients in neurocritical care. So currently what doctors have to deal with is they have to drill through somebody’s skull, insert sensors to measure things like oxygen saturation, blood flow perfusion, and intracranial pressure. And they use that data directly to make decisions for assessment, but also delivery of their therapeutics.

And there’s tons and tons of data and evidence that shows that that makes a huge difference on outcome. The problem is that this is a highly invasive surgery. There are issues such as bleeding and swelling that essentially prevent doctors from being able to drill onto too many of these patients. So they do this in about 200, maybe 300,000 patients every single year in neurocritical care.

But neurology gets way over 10 million patients a year. And so they would love to be able to drill through that data. They would love to be able to have access to this data that’s direct and quantitative if there was a non-invasive way to acquire it. And that’s what we’re doing.

You use the word drill. So that means you actually drill a hole into the skull. Absolutely. So they literally take a drill, like think about your household drill.

They drill a burr hole through your skull. And then not only that, they insert a sensor that actually goes through your brain tissue and damages some of it. Along the way, before they locate it in a spot and measure these metrics. That’s what you call invasive.

Exactly. Yes. Just saying. Okay.

And your company or your vision is to do this non-invasively. No drilling, just putting something on to your skull. We can either have little sensors that we stick onto the head and it provides this data non-invasively. Or we’ve shown that we can actually.

Perform these measurements from a distance. So for my PhD, I actually took my laser pointer and I shined it onto patients’ heads from 30 centimeters away and made all these measurements. Very cool. I’m building Zenith.

Zenith is an AI first operating system for neuropsychiatry. We can think about it as broadly like an electronic health record system, like an EHR. And so for us, our thesis is neuropsychiatry is growing. It’s going to 5X in the next five years from 4 billion to probably around 20 billion.

So we’re going to be doing a lot of research on this. And we’re going to be doing a lot of research on this. But we still find that it’s very inaccessible. It’s very expensive, especially for the people that need it.

People with PTSD, depression, anxiety, eating disorders, alcoholism. And it’s shown to be 30 to 40% more effective than traditional methods such as SSRIs, which sort of just medicate you for life. So this is really a groundbreaking treatment we found is inaccessible. And that’s really because most of the clinics are driving up the price due to high human capital costs.

So our thesis is using Zenith OS, our operating system, we can really drive down the cost of treatment through effectively augmenting administrative tasks and helping with patient recommendations as to how to best treat them with AI-based techniques. What is this treatment? For example, neuropsychiatry includes ketamine or NVMA. Probably we have FDA legislation coming this fall.

And then also psilocybin, which is mushrooms. Right. So these treatments are actual modalities to treat. These type of depressive symptoms.

And for example, today, if you go into a ketamine based clinic here in New York, you will sit down in a nice chair. It was a very nice environment. They’ll probably put in ketamine based IV drip into your vein and then go through the session, maybe a couple of hours where a therapist will monitor you. And they also will have sessions before and after just to help you get into the zone, as they say, and also help you reintegrate with society afterwards.

So you are essentially building the technology that would go alongside with that medical treatment. Exactly. Right. For example, the high cost here is typically there’s two therapists in the room because of sexual harassment.

And you’re in a very vulnerable state when you’re taking these types of treatments or these type of drugs. Right. And so that’s high cost. And so effectively, when you can even just take out one therapist who is charging, I don’t know, $100, $150 an hour.

Because they’re a psychotherapist, you can save a lot of costs and just expand it out for a number of patients you have per day. The way you can do that is, for example, you can use our AI Co-Pilot, which can help you record, transcribe, and more importantly, run analytics on patient conversations to help you better understand how to best treat them. I’m Vince Hartman, co-founder, CEO of Abstractive Health. So I’ve been in health care 15 years.

I used to work at Epic. And then I’ve been a health care consultant. I was a consultant at one point for six years. I came back to Cornell Tech, and I did my thesis on summarization of the medical chart to create a real-time summary for physicians.

I did that for two years. And in that process, I formed a partnership with Weill Cornell Medicine. And we used real patient data to build a fine-tune, a large-language model to do the clinical summarization. About two years ago, we created the company Abstractive Health.

And we have since been working on a new model. We have research pilots with New York Presbyterian. And we have pilots across the US with a number of physicians. We are in a competition with the Veteran Affairs Department, the VA, top 10 in this for AI trustworthiness.

The whole process is that we retrieve the numerous medical records through an HIE exchange that we have access to. We get them in real time. And then we summarize this content for the physician in real time using AI. And they can quickly understand their patient.

Today, a physician oftentimes doesn’t actually know who the patient is before they go into the exam room. And if they do, they only have three minutes to review the patient’s chart. And so we’re able to create that summary. And then they also, using our platform, they create their treatment note or their SOAP note using that original summary.

So the user flow essentially is you have a patient. That patient has data. And that information gets to the doctor so that the doctor is informed. Yes.

So I can give you out the full onboarding process. Yes. So we have an API direct to the National Health Exchange. Okay.

So only about 30 companies in the US have access to this. Epic and Cerner are the largest two health systems. Yes. EHR systems that health systems use.

Now, in large language models, the best way to do this is to use the language models. The big thing was attention is all you need, right? Yeah. And the big thing here was that context matters.

So if you go in as a doctor and you get the summary of 10 years of record, but in reality, the person is about to die because of a line infection, because he wasn’t active yet. And somebody drilled a hole into this person’s all I can. Then that matters most. How do you manage context here?

So we actually have a salient model structure that determines which notes are the most salient. And we’ve trained it from previous physician note context. And we use a Roberta based model. So it’s very fast and quick.

And we then determine what notes. And then specifically when we have that content. We then use Lama to specifically summarize that. And we’re able to handle a large context window.

So of about 100 to 300,000 words. Typically, if you fed all these words into like GPT-4 or I forget, anthropics like Claude’s model, the accuracy will decrease quite a lot. Once you get into larger context window sizes, we optimize this process through using that salient structure. And that’s actually like that’s our core functionality and what our patent is based on.

And the research that we did during grad school. You described the RAG model, right? So we actually don’t use RAG currently, but we have looked at. We may use RAG in the next one to two months.

So we do it asynchronously. And we have our model can do in like a batch sort of structure. The segmented notes. And then so then for each one of these segments, we determine is this important content or is this not important content based on the structure of the note.

And then so then for each one of these segments, we determine is this not important content based on the structure of the note. And then so then we determine is this not important content based on the structure of the note. what is the most important. So then it helps reduce the window.

So most medical- That is a pre-processing step. That’s like a pre-processing step, but it uses a large language model to do that pre-processing. The medical chart, even though it’s huge, there’s a lot of non-important contents. If you do the pre-processing, meaning you look through the content, you decide what is valuable and what is not valuable, that’s a decision based on a context which you had set.

So we trained it on physician notes, like a labeling structure. If the physician is a neurologist, an oncologist, an internist, they might value different insights differently. Correct. So we do not attempt to market and sell a product that is specifically to an oncologist because you just brought it up specifically that they may find value more in that needle and needle.

Yes. Like where they want to know from seven years ago, this specific blood test value was abnormal. A RAG approach may be able to handle that. Like a lot of evidence actually hasn’t been demonstrated yes or no on that.

It’s an open question in the field of gen AI, but we do not attempt to tackle that problem specifically. Got it, got it. False positive, false negative are a big topic in this field. And now you go on with, you know, the gen AI and you have hallucinations, you have way more difficulties to manage that.

How do you deal with it? We’ve done evaluative tests with physicians and the summaries are at a comparable level of like what a physician would write themselves. And so if you go back to the original problem space, the physician isn’t oftentimes creating any summary at all. They don’t have the capability to view the medical charts.

They’re not getting any useful content from it. And so an AI generated summary is oftentimes better than literally nothing. And there’s a lot of research that shows that like AI summaries when provided to a physician enhance patient care. So that’s number one.

And number two is that we do not infer medical diagnoses. So we actually, within our language model, we constrain through, it’s called beam search. That’s not a part of the language model, but it’s a part of the medical approach. But you oftentimes put it as a structure at inference to somewhat enhance the sentences so that you don’t look at just like a greedy approach you would do multiple.

So you actually, you try to avoid of giving in diagnosis because as soon as you would do this, you would become FDA approval. Here you just summarize. Exactly. So we are only in the business of condensing comprehensive medical notes.

The language is not the same. The language model is capable of if you have all these indications that you likely have sepsis or diabetes or pneumonia sepsis, there’s a lot of language models that are doing for prediction, but we, if like the medical notes never had the word sepsis, we prevent that at the time of inference. Healthcare is unique in that there’s a dictionary that the government creates for diagnostic terms via SNOMED and ICD-10. So we use that dictionary.

And then we do like a set of all words that are from our original source notes. And then we make sure that, and we also include synonyms and acronyms, that those additional words are not inferred at the time of summarization. So we integrate into the electronic health records also. It’s more just a onboarding process.

So we offer a standalone product so that the physician can like quickly get access to the data and see a real time summary. And then for organization practices, we can do EHR integration. So it’s a business. Yes, got it.

Now I understand. Yeah. Not quite, right? It’s not only a business decision, it is as well usability.

So I love the way how you use it. Now, my biggest concern is you are dependent on the workflow. Company decides workflow. You initially pitched and says we are trying to avoid doctors burning their teeth.

And they say, well, we are trying to avoid doctors burnout. Meaning, do you really have a way to charge for that value? Is there value? And if there is value, how do you capture it?

Because if there’s value, then wouldn’t the EHR try to capture it? The record retrieval component of getting a patient’s full chart, companies charge anywhere from minimum like $5 per chart, $70 per chart to get all this data. Yes. And while Epic and Cerner are part of these exchanges, the majority of daily medicine outpatient practices don’t have access to the exchanges.

And so they currently do not, they’re not even able to retrieve the records. And so if you go to a Duane Reade in New York City for like a minute clinic at CVS, actually a minute clinic is on Epic now. But if you go to any of these sort of like family medicine practices, they just don’t even have your record. And so the capability of getting it.

Has a value which we provide. It will be good for that family physician to actually have background saying, Oh, hold on. I actually see colon cancer runs in your family. You’re over 45, which you are not, but you should do a scan, which you probably should not.

But like, you know, that will be a discussion. There will be awesome for you. That person, that doctor still gets paid only $70. Yes.

So we, for the standalone product, we charge $99 per month. And so they typically see about 20 to 30 patients a day, 500 patients a month. So just from a chart retrieval perspective, it is an amazing value for that physician. Physicians get billed mainly based on medical decision making.

So if they actually understand like that, you’re a risky patients that you come with complications, that you have a number of diagnoses, they get reimbursed at a higher rate. The value creation is that you do a higher risk coding. That’s a great point. That’s a huge market, right?

Arcadia is in there. There’s a lot of companies who sometimes up code, which is illegal, but essentially using the record to better map risk profiles. Correct. Yes.

And so we offer the standalone product in a sense for them to get value, to get the record, to do the summarization. And then we have a higher tier from like a sales perspective that they get features for value based coding. For E&M level. And that’s one of the core value props that we provide.

Got it. The coding aspect. And just to be for the audience and not everybody is a European audience here in the US, you get paid based on the risk profile. So if you come and you have a certain disease and you are otherwise healthy, then it’s maybe easy for you to treat the disease.

If you are already in a state, probably not in a state of risk, then you can get paid. If you are already in a state progression or state that demands a higher level of care, it’s more work for the doctor and therefore you get set in a different risk profile. What happens very often that the doctor gets the patient with a disease does not know that this is a higher risk. Therefore, charges not sufficiently, but has all the work because it still takes on all the necessity.

So the doctor does the correct job, but gets paid less than it could be. And so you actually close that gap. That’s a huge market. Yes, exactly.

My suggestion would be to focus on that value part, which is risk coding, right? Because that’s where the money is. If you tell me patient burnout or like doctors burnout, I was like, yeah, it’s an issue. But there is no monetizable value on, right?

Which makes it difficult as a product. Right. Well, when you say risk coding, then obviously I think, well, there are a lot of companies already doing this, but you saying, oh, like my niche is family medicine. Okay, that’s a value.

That’s a workflow. And you use AI. So I will jump to the hardware part. You avoid like drilling holes in it too.

What’s your main value proposition you bring and how do you monetize it? So there’s a few ways to look at value. So who are we bringing value to? Right.

So for the patient, the value, there’s plenty of data to prove that there’s value. And even using the drill method to acquire that continuous quantitative data, there’s a lot of value in using that to assess and administer therapies in real time. Because what you’re trying to do for things like stroke or for traumatic brain injury patients, you’re trying to minimize that value. Right.

So you’re trying to minimize secondary injuries, whether that’s more bleeding or infections or edema. And edema causes more swelling, which then reduces oxyhemoglobin, which creates oxygen deficiency for neurons and you get cell death. So all these complications as a result of not optimizing blood flow to different parts of the brain. Right.

And so there’s a lot of value in that. So the patient is able to get the best care by having continuous data. The clinician. Right.

And then currently what they do is a neurologist has to come in, perform that surgery, stick the sensor in, which takes, let’s say, 10, 15 minutes, wait for fluid to build up because the device essentially measures pressure, which then is used indirectly. Hold on, 10 to 15 minutes, including drilling the hole? Yeah. I mean, it’s pretty streamlined.

Very often you see people very excited about a problem space where you could help using technology. But you cannot monetize. And I think for the audience, this is extremely important to understand that having a technology that helps doesn’t necessarily mean it’s a winner. You need to have a technology that helps and can be monetized.

And there is also technology which this can be monetized, which doesn’t even help. Even that’s better than having only a technology. Right. Only a technology that only helps.

Yes, I love it. And that you did this effort and walked through this is awesome. Value for the patient is clear, for the neurologist is clear because now they just have that data in real time without the surgery. It’s there.

We save a bunch of time, efficiency, etc. They can now bill for additional patients that they weren’t able to drill before because of contraindications. Right. So certain patients are at risk because of preexisting conditions.

Or let’s say if it’s, you know, a heart attack. Or let’s say if it’s, you know, some sort of a trauma that they’re not able to kill. Explain for the audience what’s preexisting condition. For example, patients that might be at risk for additional bleeding.

Right. Let’s say they have more plaque buildup or all sorts of diseases that are cerebrovascular that exist in the brain. They don’t want to be drilling into them. Sometimes age is also a contraindication for inserting these devices.

Elevated blood pressure could be another one. So there’s, you know, there’s a ton of these. It’s invasive. Right.

Putting a hole in. It goes fast. I did not know that this goes that fast. I’m still like struck on this one.

But yes. It works through hair. So we can either have contact sensors that they can, doctor gets to place wherever they want. They can move it around as well.

They can have multiple sensors. And we realized that all the clinicians, once we started talking to them and iterating on this product together, that they wanted about six of these, three on each hemisphere, just so they can have access to, for example, the lesion site for strokes and a non-lesion site so that they can track in real time and say, how’s recovery going? How is my therapeutic working? Is it reperfusing this region that was lacking oxygen or flow, for example?

You made it clear that this is not a device. That here is a clinical device. Yes. Meaning you need FDA approval.

So yes, there’s a lot of applications for this. Right. And there’s a reason I chose a medical device in the most regulated, rigorous space to begin with. I guess the greater field for this that you can use, that you can think of as the landscape, is human computer interfacing.

One of the bigger fields, I guess, that have trended over the past few years. And we’re almost at a bunch of different applications. The problem from if you start with those spaces, you then run into the issue of, hey, like, do we believe your technology? Because there are so many companies that essentially take off the shelf existing technology and then they apply their special sauce machine learning algorithms to analyze the data.

And then most of them have not been very successful over the past few years. So what we decided is our background is in this space of neurology where we are able to quantify blood flow in subjects and healthy subjects. We’ve done a little bit of preliminary testing in disease patients as well. We want to start with a device that is regulated, that is being used by neurologists.

And again, we have this device already in a few different hospitals. There’s a bunch of neurologists that are testing with it currently. But once it gets approval, we want to be able to say it’s something that neurologists are using. And neurologists are using to diagnose and treat and maintain strokes and brain injuries already.

And that’ll enable us to shift to other applications that are non-regulated more easily. With this device, you are helping doctors to take decisions. Yes. Okay.

That device is reducing the cost and the invasiveness of the current approach. Right. And you showed me pictures. This is not only drill.

This drilling the hole into the skull sounds terrible. Afterwards, they run around having those electrodes stick out, right? I mean, you run around if you’re lucky. And there are some cases where they’re awake.

But for most of these patients, I think that they’re under. And they’re like that for several days with that device. It’s in, right? Yes.

It stays there for sometimes 24 to 72 to longer. Now, for the usability. This is a different data set. This is not the same data as the doctor had it before.

This is essentially you just created some new data set for them. How do they know how to use it? You give them the data in the form that they’re already used to interpreting today with those invasive devices. Got it.

So you translate your signal into the data they are used to know. Yes. And then you’re measuring that exact same thing. All beat with a different technique.

Ah, got it. Okay. So in terms of machine learning, what you used to do is you actually had patients you stick the electrodes in. And you had your device.

And therefore, you could calibrate what is wrong. So we don’t use machine learning with the original data. So the original data, we’re literally using light scattering physics to quantify the motion of how red blood cells are moving inside vessels. And so we’re using physics to have real values of diffusion of red blood cells.

Got it. Okay. Right. And so we have our data looks like diffusion coefficient essentially.

It’s centimeters squared per second. It’s real values. It’s not something that’s arbitrary that you have to baseline for. This is absolute values that we’re getting.

And that is exactly what those devices are measuring as well. Oh, okay. And so now. And therefore, now you have essentially, you have the underlying.

Underlying core data to begin with. Machine learning is going to come later. I think once we start gathering more and more of these data sets, we can now start making predictions. Of, yeah.

Okay. Which is the discussion on the interface. Because, okay, so I understand the core value proposition. And I think it’s pretty awesome.

Where you lost me is on the part of that you take the hardest part. Meaning proving that a medical device is useful and you have to, there’s a lot of process. Luckily some, like we don’t want any geek coming around and saying, okay, let’s go. Trust me.

It’s just like it was an eCornell thing in the books. You don’t want that. Or like Cornell Tech thing in the books now. So it’s good that we have all this test, but this is costly.

It takes you time and so on and so forth. Why don’t you go. Directly and first prove that, for example, machine to human will work. Like fly an airplane just by having those electrodes.

Write an email by thinking about the email or whatever it is. So the contrast of what we are measuring is concentrations of hemoglobin in real time. Okay. So we’re measuring hemoglobin in real time.

Right. So oxy and deoxy hemoglobin in real time. And then on top of that, we add blood flow. Yes.

And so it’s still a step or two or three or four away from electrical activity in the brain. From being able to at least quantify it regionally. To be able to get that, we still have to do a bit of research and development with that kind of resolution specificity. And secondly, for a lot of those applications, you need coverage.

How does machine learning and data coming from the brain work? So the first thing is, you need to be able to get that. So you need to be able to get that. And then you need to be able to get that.

So you need to be able to get that. Maybe you need to be able to get motion learning and data come in? 2014, 2015, just when I had started grad school, I found a professor named Jack Gallant at UC Berkeley. And so his lab, what they did is they have state-of-the-art, super high-resolution functional MRI.

And he took his students’ subjects, put them in these MRI machines that measured the bold response, had them watch hours and hours of videos, little clips, recorded their blood flow data in their brain, created models, and then they used those models to reconstruct new videos that they were given that were unknown, obviously, using the brain activity by itself. And so you could see here’s what they’re seeing. Let’s say it’s like an elephant moving or a bird flying. And then you could see that just from the blood flow data, they can reconstruct what they’re seeing from their brain signals.

So that’s the application system. Very nice. But in order to do this, you need, essentially, it’s a classification model, right? And in order to do this, you need to have a lot of training data, which in order to connect the oxygen levels, the blood flow levels, together with the electrical activity that then leads to something people have seen.

Yes. So the way that blood flow is connected to electrical activity, it’s called neurovascular coupling. So essentially, when neurons in your brain activate, there’s this big demand for energy, ATP, and oxygen into that space. So there’s a Russian blood flow.

So that’s the hemodynamic response is what they call it. And so this coupling between the blood flow and the cardiac cycle, essentially, and being delivered to localized regions in the brain is called neurovascular coupling. So it turns out that you can use that to infer electrical activity indirectly. And it’s a very, very good idea.

And it’s a very, very slow signal. So like electrical activity on the order of, let’s say, milliseconds, hemodynamics on the order of seconds. You can actually use that to make these reconstructions still. The problem is, even with the state-of-the-art MRI systems that they have access to, so data collection is a problem, obviously, spatial resolution, they realize it’s less of a problem than they thought it would be.

Temporal resolution is where they’re lacking. So getting data with MRI, especially ASL MRI or bold MRI, you get resolution on order of seconds. So you can actually use that to figure out what’s happening in the brain. And so that’s a very, very good idea.

And I think their system was something on the order of like maybe 10 hertz. They had to make some sacrifices in other spaces, but they would love to have that data at 100 hertz, for example. And so that, in our opinion, could be possible with optics. Now, to get there, we need to be able to have hundreds of these sensors map, let’s say, the visual cortex, and then replicate what a $15 million FMRI system is measuring in real time.

And I think it’s a very, very good idea. And so getting there is equally challenging for different reasons than the medical device, right? I hear you. I like the AI vision here, because what you’re saying essentially is, so far, you’re mapping out physics, because you’re mapping out physics, you exactly know what you measure.

The challenge here is get the medical approvals done. In the other space, you say, well, if I have enough training data, because every human is probably slightly different, I’m going to have to do a lot of training data. And so I think that’s a very different, right? Because it’s not on that physical level anymore.

It is on an abstracted level. If you could do those predictive measurements, essentially, that would be cool. But you actually need way more training data. And therefore, going the medical route is for you to start to get to a setup where you can start collecting.

Yes. So that’s one of the reasons. Another reason, is to be able to have, let’s say, hundreds of these sensors or an array of these sensors, we need to have a paradigm shift in the actual manufacturing processes for this. So right now, what our device looks like, it’s a box with a bunch of off-the-shelf equipment that are customized, obviously.

And the interface is fiber optic cables that may or may not need to be in contact with that subject. That’s what we’re using for a couple of different locations. But now to get an array, the form factor isn’t what I’m talking about. So I’m talking about that black box I talked about, that has all the hardware.

Yeah. We want to put all that on a single little chip. And so, technologies for that are silicon photonics. You can actually make little circuits using light.

And we can actually miniaturize a lot of the things that we do in this space optically. We can actually miniaturize them on these platforms. And that’s because it’s already possible in areas like transceivers, optical transceivers, that convert data from analog to optical and back-and-forth, to power the internet. It’s possible in very advanced forms of lidar.

Yeah. possible in quantum computing. I think a lot of GPU hardware is now going towards that space where they want to use all optical processing if they can, at least for some of the computations. And so that field has exploded recently.

And it turns out that we can actually take our technology, use some of those processes to put it on a chip as well. And so that’s kind of a parallel effort. Could you already productionalize without having had this chip? Yes, for the medical device application of, let’s say, a few different locations, reliably and quantitatively.

I think anything that’s an out there application, let’s call it, is going to require arrays and miniaturization. And the other way that we looked at it was, we don’t necessarily have to demonstrate all those applications because there’s so many startups in this space and companies. And you can think about name any big tech company, they have internal teams trying to work through this, right? And so our goal is if we can actually manufacture this on these chips, we can be the supplier of these chips and hardware to all these companies building these applications.

So you have this bigger vision, which is essentially the chip allowing you to measure electrical activity by knowing what to infer from that physical activity. So that’s one way. It turns out that if you make it really, sensitive by, let’s say, a million times more-ish, which is actually possible on some estimates, you can measure direct electrical activity as well. I think that there’s paths for both, whether it’s direct or indirect.

Perfect. Awesome. You can do it from a meter distance. That would be, like, this is pretty amazing.

And because it’s on a small chip, you could put it on a device like an iPhone. In order to get there, you first start with the medical application. A hard road to get started, but at least you have a data set. For the medical application, are you just measuring the, like, quantitative values and presenting them to the physician?

So you’re not, like, saying patient has stroke, no stroke. So it’s a slightly easier path of FDA approval to just do the scores and the values and the actual medical prediction. They’re using it for medical diagnoses. You’re not actually doing.

The physician is still doing the medical decision-making. The pathway is still going to be a de novo approval process, which just means they have to classify the device, and then the clinical study has to be a bit more rigorous. But we still expect this to be a class 2, maybe even a class 1 device, because it’s completely non-invasive. And then the evidence required is to co-register those invasive devices with our device, and we’re actually in the process of doing that.

We’re in the process of setting this up. Oh, okay. And saying, look, it gives you the exact same answer over the course of, like, a week, for example, that a patient’s been wearing this. And therefore it becomes easier for you to go to market.

And then you say, my accuracy is within a few percent of the invasive device. I give it to you in the same format for how you interpret the data, and that’s how you show that it’s. Very, very cool.