Become a Creator today!Start creating today - Share your story with the world!
Start for free
00:00:00
00:00:01
India's Non-Invasive Answer to Elon Musk's Neuralink
1 Playsin 10 minutes
Most Indian AI startups are racing to build chatbots for Indic languages. Dr. Siddharth Panwar took a different bet, building India's first non-invasive brain-computer interface on 60,000 hours of EEG data that hospitals were about to throw away.
Trained at Stanford as an electrical engineer and forged at IIT Delhi as a brain scientist, Dr. Siddharth Panwar spent ten years sitting outside neurologists' rooms collecting EEG data that nobody else wanted. That patience became NeuroDx.ai, the deeptech startup behind MANAS-1, India's first 400-million-parameter brain foundation model trained on 60,000 hours of EEG signals from over 25,000 patients.
In conversation with host Akshay Datt, Siddharth explains why the West gave up on EEG too soon, why a non-invasive brain-computer interface can capture 80 percent of Neuralink's value at a fraction of the cost, and why Indian VCs still struggle to underwrite frontier scientific risk in AI neurodiagnostics. Selected as one of twelve IndiaAI Mission sovereign AI champions, NeuroDx is the only physiological foundation model in the cohort, arriving exactly as Indian deeptech enters its sovereign AI moment.
👉How Siddharth Panwar built India's first 400-million-parameter brain foundation model, MANAS-1, with zero institutional VC on the cap table
👉Why the Western medical system gave up on EEG and how AI foundation models are bringing it back into clinical relevance
👉What separates MANAS-1 from Neuralink, and why a non-invasive brain-computer interface scales to populations that brain surgery never can
👉How NeuroDx aggregated 60,000 hours of EEG data from 25,000 patients by salvaging signals that hospitals were deleting every month
👉Why being selected as 1 of 12 IndiaAI Mission sovereign AI champions changes the financial and policy economics of building deeptech in India
Subscribe to Founder Thesis for weekly founder conversations and follow Akshay Datt on LinkedIn [https://www.linkedin.com/in/akshay-datt] for daily insights.
00:00 - Introduction
00:02:14 - What Brain Computer Interface Really Means
00:03:09 - Non-Invasive Alternative to Neuralink
00:04:11 - EEG: India's Forgotten Diagnostic Edge
00:13:36 - EEG as the Language of Brain
00:16:19 - MANAS-1: India's Brain Foundation Model
00:21:02 - 400 Million Parameters Explained Simply
00:30:09 - Inside the IndiaAI Mission Selection
00:35:00 - Bootstrapping a Frontier AI Lab
00:49:19 - Why $10M is India's Biggest Test
#NeuroDx #SiddharthPanwar #FounderThesis #AkshayDatt #MANAS1 #BrainComputerInterface #BCI #IndiaAIMission #DeeptechIndia #IndianStartups #Neuralink #EEG #FoundationModels #SovereignAI #AINeurodiagnostics #NonInvasiveBCI #IITMandi #BootstrappedStartup #HealthcareAI #IndianDeeptech
Disclaimer: The views expressed are those of the speaker, not necessarily the channel
Recommended
Transcript
Introduction to NeuroDX and EEG-based foundation model
00:00:00
Speaker
Dr. Siddharth Panwar is building one of the most ambitious AI companies from India. He's building a foundation model trained on EEG signals, which will help in diagnosing neurological disorders. Eventually, it will be able to read your mind.
00:00:10
Speaker
dr siapanor is building one of the most ambitious ai companies from india he's building a foundation model trained on e easy signals which will help in diagnosing neurological disorders eventually it'll be able to read your mind Siddharth, welcome to the Founder Thesis podcast. You are the founder of NeuroDX. except My first question to you, but what what exactly does NeuroDX mean?
00:00:33
Speaker
Yeah, so NeuroDX, ah the neuro stands for neurology and DX in, you know, typical medical parlance stands for diagnosis. So ah the term together means neurodiagnosis. Although our company aims to do much bigger things and beyond neurodiagnosis, but this is where we are starting.
00:00:50
Speaker
So you are operating in the domain of brain computer interface, BCI. ah What does that mean? what What exactly is that domain? So ah getting the technology around us to be able to talk to the brain directly in terms of providing an interface, a direct interface to the brain. Right now, the interface to technology is through our fingers, through our mouth, speech, right?
00:01:15
Speaker
But if you can provide a direct interface to the brain ah with the ah any kind of sensor that you can put in the brain or on top of the brain, That will be typically called brain-computer interface where it can then start doing things where you just have to think about them.
00:01:32
Speaker
And our model is essentially decoding the brain and on top of it, we'll build an interface so that you have access to
Non-invasive BCI technology vs. Neuralink
00:01:40
Speaker
it. This is what Elon Musk is doing at Neuralink.
00:01:43
Speaker
Yes, he's doing it invasively. We are trying to do it non-invasively. So, invisibly means Neuralink is cutting open and placing a chip inside. Yes. yes They do a surgery.
00:01:55
Speaker
Right. Yeah. They do a surgery, which is ah something that takes about a day, I think. And they put ah a small chip in a very specific location of the brain. So, it does a very specific task. ah If it's trying to decode speech, they'll put a chip where the speech center of the brain is.
00:02:13
Speaker
If they're trying to maybe decode what the vision is, how the person is seeing and what they are seeing, then they'll put a chip in the vision center of the brain, which is at the back of the head. So it's a very local, very specific ah intervention. But the advantage is that it is able to decode signals and neurons activity very precisely because it's sticking right to the neurons.
00:02:32
Speaker
It's basically getting high res signal. Yeah, exactly. High res.
History and applications of EEG technology
00:02:36
Speaker
What you envisage is based on the EEG technology. Ah, okay. EG stands for electroencephalogram.
00:02:45
Speaker
ah Electro is the electrical activity and cephalo is the brain part. And the gram is the measurement, right? So ah we basically pick up the electrical activity of the brain, which is happening because neurons are, sorry, the ions are moving between neurons to communicate. with each And that electrical activity goes out and through the scalp can be read.
00:03:06
Speaker
So we record that and by that we try to understand how the brain is functioning. So EEG stands apart from MRI and CT because um MRI and CT take pictures of the brain. They see how the brain is structurally.
00:03:18
Speaker
EEG looks at the brain in terms of how it's functioning because it's looking at the electrical activity. So once you put a number of electrodes on the scale, Typically, it's around 20 to 30 in a clinic.
00:03:30
Speaker
Then that seen in know you know a piece of paper as time series, that is called an electroencephalogram. That's an H. And with that, we are hoping to decode the brain.
00:03:42
Speaker
This time series has 20, like if there are 20 electrodes, ah there'll be like 20 lines moving up down. Yeah. Yeah, or they're combinations, right? So you can make more combinations. And so those things are obviously variable.
00:03:54
Speaker
You can have a single electrode EEG also. You just put one and. How old is the EEG as a technology and what is it used for? It's the first technology for the brain. It's decades old. it's It came before anything came around.
00:04:07
Speaker
Hans Bercher is the one who actually came up with this technology a long time ago. And it was the main modality to diagnose the brain. And back in the 60s and the 70s, the people using these technologies were very, very skilled.
00:04:20
Speaker
They could look at an EEG and tell many things. ah Right now, its domain has become very restricted because of the advent of CT and um MRI. But let's say there was a tumor in the brain.
00:04:31
Speaker
The skilled EEG technicians could tell a lot just by looking at the EEG. So that skill, because that's not passed over in a textbook, it's like how well can you read a signal, right? So it's not very precise. So therefore, it's a tri it was tribal knowledge. It's something that was passed down from a person to person.
00:04:48
Speaker
So that got lost. And now EEG is in the clinic. Every hospital has it, by the way. But it's used in a very narrow domain, very specific things. What is it currently used for?
00:05:00
Speaker
It's currently used for ah diagnosing epilepsy. very precisely epilepsy. It is also used in some auxiliary things like sleep studies and it's also used in maybe sometimes in dementia. But there it is kind of a supporting thing. It's not the main modality. But in epilepsy, it can become the main modality. But even there, it has a lot of ah shortcomings and disadvantages, which we are trying to overcome as a first offering.
Limitations of EEG in epilepsy diagnosis
00:05:28
Speaker
Okay, what are the shortcomings? like It's because the tribal knowledge is not there. Is that it? like That's one. But to in epilepsy, the advantage is that the marker of the disease is very, very obvious.
00:05:41
Speaker
It doesn't require any ah tribal knowledge. It's a very specific shape. And these yeah shapes, there are set of these shapes and they have been codified very precisely in medical literature.
00:05:51
Speaker
So they can be taught to a student. Here is a shape. If this appears, you have to say this is epilepsy. Because of that, EEG is still relevant in epilepsy. But unfortunately, and here's the shock coming, let's say somebody has epilepsy and they take a 30-minute EEG recording, which is the standard across the board.
00:06:10
Speaker
It is not necessary that that shape will appear in the EEG. It's a statistical phenomenon. It might not appear. And if it doesn't appear, how do you diagnose them? right If a bone is broken, it's broken. You take a picture of it, it's broken. right If there is a patch on you know on the lungs, it's there. It's always there.
00:06:29
Speaker
But how do you find something in EG which is always there is the task neurodegesis. Those things are not visible by the eye. They are visible to the model. But they are always there.
00:06:40
Speaker
That's the premise of our work. Okay, got it. So EEG, I'm assuming, would be significantly cheaper than a CT scan and MRI, which are the other alternatives to look inside a brain?
00:06:52
Speaker
Yes. In Indian machine costs, it's cheaper in all kinds of ways. So first is obtaining the box, which does EEG. It's a very small box. So it's cheaper in its spatial yeah investment.
00:07:05
Speaker
It's cheaper in terms of money. it's a store In India, you can get it for the three lakh rupees, right? West, it's more expensive. It's cheaper in terms of what it takes to run the machine. A person who's ah proficient in EEG or who needs to be around for taking an EEG is somebody who doesn't need to be an MD or a radiologist or an MEBS.
00:07:23
Speaker
Somebody who's just done a diploma, right? It's cheaper to run the machine. if the electrodes and everything, you know, they can go. and And deployment. let those cost server Yeah, it's very, very cheap, right? Deployment, like bringing an MRI, a big machine often requires tearing down the place and putting the machine there and then building it up.
00:07:44
Speaker
right So it's ah very, very handy and very, very cheap. Again, and and so and the data is also, you know, it's not as heavy a data. Imaging data is very heavy.
00:07:55
Speaker
ah This EEG data can be easily sent over through the internet from a, let's say, remote location on a cloud and we can process it. So it's an incredible tool which has, as I said, the West has completely ignored for the longest time, but it won't for too long.
00:08:10
Speaker
It's going to do a comeback. So you're saying in the West, even for epilepsy, they rarely use EEG. They straight away go for a CT scan or an MRI. Yeah, that's what I'm seeing, which is shocking to me. I didn't believe that statement you just made. I thought even in epilepsy, at least it's being used.
00:08:26
Speaker
But repeatedly, I have seen things from my friends who have taken asper opinions on things that were happening in the US. So I lived for 10 years in the US. And what I'm seeing is completely shocking to me.
00:08:38
Speaker
In some cases, they are not even doing an EEG because that skill to even look and the willingness to even look at something which is not as precise, which may be a little bit of subjectivities is involved, it just goes out of the window. Partly it has to do, maybe my guess is because it's a very litigious society.
00:08:55
Speaker
The medical legal framework is very, very rigid. So if they make a decision and it cannot be backed very rigorously by a test, you know, that becomes a problem for them. And moreover, I've seen assessments done by people, reports that I obtained by looking at the EEG, what reports were generated.
00:09:12
Speaker
And I found them a very subpar. I mean, even i as a non-EEG person, i mean, I'm just gotten used to seeing EEGs. I could see that that expert had done mistakes in the report.
00:09:25
Speaker
basic mistakes. So it just goes to show that they have to stop using it. They just don't rely on it. right Although they have it in their clinics. So yeah, um that's unfortunate.
AI models decoding EEG data for better diagnoses
00:09:35
Speaker
But it's an opportunity for us. So now the thing which AI makes possible is the ability to decode the EEG lines. So like you said it's a time series. It's a graph, basically. AI makes it possible to decode that graph and give predictions or give give diagnosis, ah which is what NeuroDX fundamental thesis at this stage is. Yes, the first one, yeah. So BCI is a broader term because BCI is claiming to be able to read everything that is going on with the brain, whether it's clinical, like neuroling, right? So it's going from clinical to hopefully one day commercial and more applied things that are affecting ordinary people.
00:10:18
Speaker
But the first a place where you put your wedges, clinical stuff. So ah diagnosis happens by reading things that, for us, but reading things in the data, which a human eye is just simply not capable of reading. And it's plenty of information is there.
00:10:33
Speaker
So how does and ah how do you build an AI which can read graphs? ah Is this like a large language model that you're building or but what is it that you're building? Yeah, essentially, right. So in the beginning, so it was a journey, even a technological journey. So when I got to started on this problem, this was um like 2012 and AI hadn't taken off.
00:10:58
Speaker
So what I was doing to do solve this problem was apply the the standard techniques of those times, which was deterministic algorithmic approaches. which is broadly called signal processing.
00:11:08
Speaker
You take the data, you process it in some mathematical fashion and hope to pull out some features from it which you can then use to diagnose. Then came the era of narrow AI, right? And so what you would do is you collect two sets of data. This has epilepsy, this doesn't have epilepsy and build models, supervise learned models that try to identify them.
00:11:28
Speaker
And it does does a better job than the previous version in some ways. ah But then it has its own shortcomings and how it is able to generalize across patients, across different types of machines and other of that.
00:11:39
Speaker
But then comes along foundation models, right? And so I treat EEG as the language of the brain. It's no different from, you know, it's sequence, it's speech. is When I'm speaking, it's a time series that you are hearing.
00:11:53
Speaker
And inside it is embedded a language, right? Your brain decodes that language. And then a chat GPT goes and breaks that apart, puts some attention back and forth, sees the relations between, success you know, a sequence of essentially air pressure.
00:12:10
Speaker
What I'm saying is air pressure. Right? Yeah. So ah I don't see EGA as anything other than a language because it's a sequence. It's very similar to speech. And in fact, there are 20 people there. So it's like a chorus.
00:12:24
Speaker
I see it as a chorus, 20 people singing. right So you can build models using almost identical approach ah as these large language models, which is what we are doing. So our first offering essentially used the classical techniques ah to decode the brain by existing transformer type methods.
00:12:44
Speaker
So you take a lot of EEG Just like ah these language model companies took a lot of text data, you throw it at the model and ask it to learn it. right You basically hide portions of those EEGs and say reconstruct it.
00:12:57
Speaker
Just like you hide some portions of the language and say reconstruct it. So the model eventually starts figuring out how to reconstruct EEG. So it basically understands how EEG is eg constructed in the first place.
00:13:10
Speaker
And so if you get enough data, it gets all the kind of variety that is around EEG and how to construct it. And then you can start fine-tuning it for your specific tasks. You want to do epilepsy diagnosis, you want to do dementia diagnosis, you want to do anxiety, day whatever. And then you go to BCI applications.
00:13:27
Speaker
You are thinking about something. ah You are feeling an emotion. You're frustrated, whatever. right So you can just kind keep building on top of that foundation. a specific architecture, a specific you know application that suits you the most.
00:13:43
Speaker
And now the advantage of this foundation model approach is now that you need much less data than you used to ah before to build these kind of applications. And they are far more robust as well.
00:13:55
Speaker
With less data, more robustness has come because you have a foundational layer on which you are building now. So exactly like essentially exactly like a ah language model, but we are now in our next revisions are trying to bring in different architectures which are more relevant to the EEG because EEG is not just in time, it's in space also.
00:14:18
Speaker
It's a spatio-temporal data. Language is a temporal data. It's just evolving in time. So there are things that you can modify in that same architecture, even get some newer ideas, which our next release hopefully is going to talk about, Manas, second, Manas 2.
00:14:35
Speaker
ah Manas is the name of our model. I want to um go one by one. So you said that ah your approach first was deterministic algorithms. I am...
00:14:48
Speaker
interpreting this, you tell me if my interpretation is right or wrong. So a deterministic algorithm is where somebody has applied their mind and created a formula if then. Yeah.
00:15:01
Speaker
Yes. So so this that is a human engineered formula for that that if the peak is so high or if the angle. Yeah. I can refine it a little bit.
00:15:13
Speaker
Yeah, I can, gri up just you are on the money. ah So what you do is ah the but the knowledge ah and information is derived from following deterministic pathways.
00:15:25
Speaker
AI is basically learning about information in a statistical manner. You you know, you it go goes through data and it's essentially learning something which is average of what is being presented to us. This is how where our brain works, by the way. So it's mimicking the brain.
00:15:39
Speaker
But deterministic is more like you you're not relying on ah learning the patterns. You're looking directly at it. Even if there was one data point, you know exactly where to look and then you pull it out. And that's what you use for doing other inferences. in the all Right. Okay. And then narrow AI you have... ah ah labeled and unlabeled data you said ah the narrow app is basically like machine learning yes yes I mean a classical machine learning so you have a very narrow task that you have given the machine to do and you have shown that narrow tasks in as much elaborate manner as you can
00:16:16
Speaker
That would be like this EEG belongs to an epilepsy patient and it has 500 EEGs of which 200 are of epilepsy patients. So it's able to analyze a ah new EEG and then make a prediction that this is probably epilepsy patient. Yeah. Yeah. Yeah. Okay.
00:16:36
Speaker
Okay. So it's like, let's say, an alien walks to to on this on this planet and you tell them that we speak, we communicate using this language. This has never heard this language. And then you tell it that you have to tell whether this person who's saying is angry or is he happy.
00:16:52
Speaker
So you give him a lot of angry speech. He has no idea about the language or the speech at all. But you give him some and idea of ah angry. You give him some idea of, you know, happy. And he'll start picking something or the other from it and be able to tell.
00:17:06
Speaker
Right? As opposed to somebody who becomes native in the in the speech, which is what foundation models do. You learn the language and then maybe if the language even changes, because you have learned enough foundational things, patterns that are across human language, you go to an Italian and that guy is angry, you'll be able to pick up, oh, this guy is angry. I can sense, I don't know the language, but he's angry because you're understanding aspects of speech in it or the ordering of how the words are put together, something, right? That's what foundation models do.
00:17:35
Speaker
Got it. Okay. So foundation models are basically like ah consuming vast amounts of data without labeling. You don't tell a foundation model that this EEG belongs to an epilepsy patient. You just give it a vast amount of EEG data and on its own, ah it creates some patterns. Like what does it do on it?
00:17:57
Speaker
it learns it understands how the EEG was how the EEG actually gets constructed so it's like learning a language essentially and then you learn to do specific tasks okay like a large language model will learn he is and not he are for example because it has enough patterns Yes, yes. ah It learns the structure of the language itself. The sequence has a structure, inherent grammar, right?
00:18:22
Speaker
You can talk about ah in our grammar, how things are put together. It will learn that. How are things putting being put together? It will learn the easiness of the data.
00:18:34
Speaker
Okay, got it. Okay. so So you created that foundation model, which is minus one. Yeah. Now, minus one is a 400 million parameter model? Yes, yes. What does that mean, 400 million parameters? ah So, um I mean, I'll try to explain it in terms of something very basic. ah So it's it's the knobs that you have to be able to move around and fit something that you're observing to a model. So model is a replication of what's happening in the universe, right?
00:19:08
Speaker
And ah you can... Try to build a model which is very simple by having just one knob and tweaking that knob and try to fit what you're observing to your model because that knob will then start to replicate what you're seeing, right?
00:19:22
Speaker
So these 400 million parameters are nothing but knobs that you can independently move to fit and learn something that is happening in the nature, right? So as they say, all models are wrong, some models are useful. There is no model that can be fully accurate because it's a model. It's a replication or a sort of imitation of something happening in the natural world.
00:19:44
Speaker
And the closer you get to it, the better it is. Right. And so we all try to get closer to it. And the easiest way to get closer to it is to increase the number of knobs under the architecture of your modeling that you have you know committed to. So there are ways in which you can arrange those knobs related to each other, related the data and then start making it.
00:20:04
Speaker
If you completely restructure the way that knobs are arranged, then you have a different architecture. And so if certain architectures are very finely tuned to a specific type of data, you might get away with fewer knobs because it's inherently leveraging the data ah structure. right So this is the way. What's ah like say a Gemini's parameter count? At current, I've lost track of it. It must be, I mean, these things, language models, now I've entered the billion era.
00:20:32
Speaker
Double-digit billion or even triple-digit? Yeah, yeah. Double-digit billion, it would, yeah. But you don't need a billion parameter model because EEG fundamentally is... ah More restricted and like language. So ah ah the thing is, ah we would we are actually, to answer your question, we are headed that way. We are building billion metaululate models.
00:20:52
Speaker
What that first requires is ah data of that proportion, which nobody in the world has, except us, at least partially. ah So we are headed that way, right? So if you go on the internet, you can scrap data and that's like 25,000 people.
00:21:09
Speaker
This is not enough to even try. It will be foolish to build something around billion model parameters, right? So if the data is not enough, the idea is that adding more knobs is not going to help you because you don't have richness enough richness in the data to have the complexity in the knobs and the number of knobs you have or the parameters have.
00:21:29
Speaker
So the first ah thing to do is to see how far you can go. I keep telling my engineers that we have to squeeze water out of a stone as far as EEG is concerned. If somebody is able to figure out something or to where two months from now, three years from now that we haven't figured out ourselves, then we haven't done our job.
00:21:48
Speaker
That's I'm very clear about it. So we have to push the frontier as an EEG, AI company. And for that, we have to take the models to that limit. um The data is, after all, coming from the
Challenges in EEG data collection for AI models
00:22:00
Speaker
brain.
00:22:00
Speaker
It's the most complicated piece of, ah you know, it's the most complicated entity in the known universe, the brain is, right? And so we have to treat it with respect.
00:22:12
Speaker
I've learned through my PhD that you have to treat brain with respect. So we have pushing the boundaries ah in terms of the the sheer ah ah you know size of the model. What we need...
00:22:24
Speaker
the data to go along with it. So while your assessment is correct, in some ways, if you have a little bit of data and the tasks are ah fairly easy to do in terms of, let's say, um some ah clinical applications where the data is very clearly ah obtainable by the modeling that you do. It's not hidden too deep into the data, right?
00:22:48
Speaker
then you can manage. But what we are aiming for is full-blown BCI in some sense. We want to be able to decode speech. For which you need a billion kilometer. Want to go there? Yeah, just just like OpenAI kept pushing the boundaries till they realized something emerged, right? So the the like ah the the models that these guys have built is an emergent phenomenon.
00:23:08
Speaker
This was not something that they were expecting or like ah they knew for sure. They kept pushing the boundaries and suddenly understanding and knowledge emerged. So we are on that same trajectory and we want to see how far it goes.
00:23:22
Speaker
For that media data, of course. This 400 million minus one model was built on how many EEGs? ah We ah b built it on the same number of EGs. So we wanted to benchmark. So the the world currently has 25,000 EGs. We have much more EG than that. We built it on the same 25,000 model and we did ah we are the best performing model right now.
00:23:42
Speaker
So we just wanted to you know tell the world that we can do the best job on the same amount of data. And now we are ready to take off, right? Adding more data, i'm adding more architectural improvements. So To answer your question, 25,000 EEG is what we have. And minus two is going to be on how many EEGs? Uh, minus two will definitely push a hundred thousand EEGs. Definitely push.
00:24:08
Speaker
And minus one is an open source model. Yes. What does that mean for a model to be open source? Yeah, so okay so yeah it just can be fully replicated. So if you want, ah so the the algorithm has been made given to you and you can therefore, if you can get the same data, so let's say there is a lot of data on the web available, right?
00:24:28
Speaker
And all the architectural changes and the way we have learned the algorithm of learning, we have made available to the people. You can go out, just take that algorithm, throw it on the data and you will have the same model as Beatrix.
00:24:42
Speaker
So open source is also called open weights. Like the, ah what is weight? Open weights? Yeah, yeah. They are related terms. They are related terms. What does weight mean here? Weight of what?
00:24:53
Speaker
Weight of the knobs? These 400 million knobs? Yes, yes. Exactly. Exactly. so So the model is the product of data marrying algorithm. And then the model appears in terms of exact values of the weights that need to be there.
00:25:07
Speaker
I can give you the exact weights also, which you can use directly without having to go through the process of learning and to and know and know and ah investing in the whole learning paradigm.
00:25:18
Speaker
So it's a quicker solution for you as far as being able to build on top of it. So we have done open-sources, open-weights as well. So it's an open-source, open-weights model. Why would... the anyone invest in training a model and make it open source? but Because that that means that people don't have to pay you money to use that learning, to use that algorithm, to derive at those predictions.
00:25:44
Speaker
That's a good question. That's a good question. um Well, one thing is, ah so you' you're essentially asking a question where it makes financial sense, right? So the answer is yes, of course it does.
00:25:55
Speaker
What we are hoping to do is open a new market altogether. It's ah opening up these ah brain interface-based ah offerings is not something that we want to just hold up to ourselves. We can't do everything that is possible with the BCI, right?
00:26:14
Speaker
We'll always be limited in the terms of applications that we can build on top of. And so we want people to use it. We want people to, you know, so we will make these models available. We will also make a 20 billion model available to the people.
00:26:27
Speaker
But obviously improvements will keep happening and if you want the ah you know the kind of ah model that is at the very edge at any given time, then hopefully once people start using it and building applications on top of it, they can then start talking to us in terms of making the very best product that they want to make. right So our goal is to have a student studying in at a college in India, let's say, thinking that maybe I can build, you know, a tool, a quick tool where I can control my remote control car with my thoughts, let's say.
00:27:03
Speaker
Right. And so instead of just going for a whole set of EG, just downloads a model, records new data points and, oh, it starts working. And then maybe things are building a a company or a small, you know, startup around it.
00:27:15
Speaker
And so this is what we are kind of hoping to build. Moreover, it's a commitment that we have made to the government of India because we are also part of ah a team of 12 companies which have been selected by the government to build sovereign models. um So it's a commitment we have made. Under the India AI mission. Yeah, under the India AI mission. Yeah. yeah ah Is it right to say this would be like, say, Gmail? ah Google gives Gmail a free, but they charge companies for... Yeah, enterprise versions. Yeah, there's an enterprise version of Gmail. Yes, yes, yes. So so there's something similar in terms of the strategy. Yeah, yeah, yeah. Got it. Okay. yeah And I will this really really believe that you the more you spread, the finances kind of take care of itself. You give people things to use, money takes care of itself, right? It always comes back.
00:28:05
Speaker
Okay, so you said ah that you built and modeled 400 million and then you did fine tuning. So what do you mean by fine tuning?
00:28:17
Speaker
Yeah, so what do you do then is, let's say we are, so one of the applications that are being built on these foundation models that we are owning do ah is ah is clinical applications. So just like ah OpenAI built Sora and they tried to own it for video generation. They have shut it down, by the way, because nobody was using it anymore.
00:28:35
Speaker
But that's what they hope to build, right? So similarly, what we are trying to do is we are building these ah medical applications and offer it to the doctors. So now this foundation model needs to be made very, ah you know, precise in terms of solving that particular problem, which is can you tell whether this person is epileptic or not?
00:28:58
Speaker
And so now you kind of give it the same supervised problem, which I was talking about when we're talking about narrow AI. So now instead of starting from zero, you're starting from a foundation.
00:29:09
Speaker
And you're telling this model, here is some epileptic people, here are some non-epileptic people. And now you tweak your your knobs a little bit to be able to do this job precisely.
00:29:20
Speaker
Okay. So making it fit for one specific use, in a way like how a child at 18 will join a college and learn whatever, say becoming a dentist. So that is in a way fine-tuning of his knowledge. It's a specialization.
00:29:36
Speaker
Right, right, right, right. Okay. Okay. So you are fine-tuning your model primarily for epilepsy? That's where we start. ah We will ah eventually cover the whole range
Expanding diagnostic capabilities with AI
00:29:47
Speaker
of neurological disorders. Within two years, we will have a brain scan for, like what we call brain scan. We have some offering for every age group.
00:29:54
Speaker
So ADHD, autism for younger kids. I mean, epilepsy affects all age groups. Depression, anxiety is affecting people in the middle age group. And later on dementia, which is going to be our next, we are ah putting the next most effort in dementia. So our product or a screening for everybody about the age of 55 years or above is something that we are working on so that you can just put a headset for five minutes and figure out whether there is some cognitive decline that has begun to set it. And if it day if you find it, then you can go and talk to a neurologist.
00:30:27
Speaker
ah How are some of these things currently diagnosed? Like how is ADHD diagnosed? ah ADHD I think is mostly a behaviorally diagnosed so a lot of so even dementia let's say ah because there is no as I said you ah it's not easy to just take an MRI even if it was feasibly we just not take an MRI right right off the bat So you first do some ah kind of behavioral tests and some cognitive analysis to a certain whether this person is having some you know problems in, let's say, focusing, remembering, all those things. And so that's where things start.
00:31:07
Speaker
ah at at first. But that also requires experts and it brings in subjectivity also, right? So what we are trying to do is eliminate all of that. So you're saying for diagnosing a dementia, typical process is first behavioral diagnosis and then an MRI because MRI is expensive and you don't Yeah, in an advanced stage. Yeah. So a dementia becomes completely apparent in a much advanced stage. Earlier onset, you might not be able to see it in a PET exam or ah whatever, and MRI, or whatever way of you are trying to look into the brain.
00:31:39
Speaker
Uh... In India, the gold standard for dementia is MDG PET, I believe. It's a PET scan. What is a full form of PET? It's tomography. It's a type of tomography. Ozytron, I forget the name, but it's a type of tomography.
00:31:52
Speaker
So, ah it's ah again, it's a way of looking into the brain, essentially. um We don't work in imaging modalities. But the way it works is that if something kind of shows up in ah a DG PET, for example, you will start seeing changes happening which are correlating with dementia. That is considered currently the gold standard of diagnosing dementia. Which is expensive, I'm assuming. It's not easily available, even in India. yeah Very few hospitals would carry a PET scanner. Yeah, yeah, yeah. And this debate specific flavor of a pet, FDG.
00:32:29
Speaker
Okay, okay. but I'm not sure if everybody, every, but you know, I really haven't looked into, but but I have heard through conversations and talking to people, it's not something that is easily available. like so So mostly right now it's behavioral diagnosis, which you are saying ah through an EEG, which is fed into Manus 2, for example, or whatever version of Manus.
00:32:50
Speaker
Even Manasuan, Manasuan is fully capable of doing it. We just have to collect fine-tuning data now, which is also our speciality, right? So our company's speciality is getting hold of this data. I've been working on this for 10 years. So last 10 years, I've been actually collecting data.
00:33:04
Speaker
So I can imagine you would have data of epilepsy because and and that's how EEG is used for epilepsy detection, but nobody would be able to give you label data for dementia, for ADHD, for depression. How will you collect that data? We have to do it ourselves. We have to invest in it ourselves.
00:33:24
Speaker
And a that you do by, you know, ah putting in the effort. Like you have your EEG machines, you go into the hospital, you spend the time in terms of years and months collecting it over and over again, right establishing those labors that this person actually has, ah let's say ADHD.
00:33:42
Speaker
That kind of work has to be done. Nobody wants to, you know, this is where taking the bull by the horns comes. You're bidding new guests. So that kind of effort has to be there. We're not automating human effort. If you were just automating human effort, it was just like, go get data, build models, and start setting.
00:33:58
Speaker
Right? ah That's easy to do. But then you have 100 more people who can do it. What we are trying to do is something that requires a lot of effort in terms of even collecting data, modeling data, everything.
00:34:09
Speaker
Right. So ah this is a ah very ah unique effort. looks say So your moat, your secret sauce is this data. Data. Okay, ah give me some hints on how you're doing it, because you've not raised much money. I'm just struggling to understand how a bootstrapped company can spend and data, because this sounds expensive, right?
00:34:36
Speaker
Doing EEGs. It is not actually, right? So what happens is that... ah What happens is when you ah are working in a domain where people absolutely don't fruit care much in terms of the value that it carries, that money, you know, that kind of financial implication doesn't appear. So there are... And I mean, now I think...
00:34:57
Speaker
The cat is almost out of the bag. But we have worked with a lot of collaborators over the years. our arbam um you know ah the In India, for example, there are a lot of these centers where EEG g is being done.
00:35:09
Speaker
Right. Okay. Like a diagnostic center. Yeah. Right. And so ah without going into the numbers of it, we ah you know you can actually engage with them and get good amount of data.
00:35:21
Speaker
I mean, they would have like thousands of EEG scans, which would be labeled also. Yeah. And they delete it every every month, right? It's just getting washed out of their hard disks. It's like just a radhi, like radhi, right? Our papers we used to sell. Right?
00:35:37
Speaker
So you can buy it radhi kebab, as they say. Right? But that wouldn't give you for new tests. That'll give you for epilepsy. So for building foundation models, you can get that. For these things, you have to actually go out and do ah these studies in hospitals and you have to give it time, right?
00:35:57
Speaker
Carrying off these studies is not that expensive, right? With an EEG machine, you can get an EEG machine. But you bootstrapped and pre-revenue, right? You've not made a single group.
00:36:08
Speaker
So how are you funding all this? Oh, we have ah people ah who are committed to it. So we have, ah ah so it's there is an investment in the sense that there is a company called Neuron in Bangalore.
00:36:20
Speaker
ah These are ah people and it's basically run by two people, Rajesh Kamra and Sandeep Singh. They saw what we are doing early on and they came on board and gave us the resources to kind of do what we are doing.
00:36:34
Speaker
For equity or for free? For equity. For equity. Got it. Okay. Okay. Okay. So just like Microsoft and OpenAI deal, where Microsoft gave compute to OpenAI. Yes. you Compute, anyone, financial support, engineering support, right? So you can think of it as something like a friends and family round, but more than that, right?
00:36:53
Speaker
So Sandeep is now our full-time CTO, by the way. So he ah has completely kind of gotten into this mission with us. And so we got a lot of support from them, engineering support. They already have an office. Our engineers are sitting in that same office.
00:37:09
Speaker
So it kind of like truly bootstrapping. This is bootstrapping.
00:37:16
Speaker
Yeah, so you'll be surprised how much you can, if you can trade off time with money. we As I said, I've been doing this for 10 years. I have been sitting, I used to sit for five hours in front of a doctor's hospital room. Five hours, just waiting for him to be done with his patients.
00:37:30
Speaker
I have put in the time where I didn't have the money. Or I could have just called, I want this much data, I have this much money. But I gave the time. Right. And so I built relationships. I talked to people. You know, I earned their trust. And I kind of, people who are willing to see the power of it, they help.
00:37:46
Speaker
You'll be surprised how much help you can get if you are just be willing to be patient. And you if there is a missionary zeal inside you, there are people who connect with it. If you can see something is going to happen.
00:37:56
Speaker
my My podcast is proof of that, right? When I started, I didn't have anything, but people came and they introduced and opened other doors for me and people kept coming and kept coming until I reached where I am today. yeah I think a it's it's just about um a positive mindset. like I have now finally committed to an approach. we are digressing here, but I have finally... kind of committed to my instinct that I will go with it with a specific instinct with a very positive mindset in terms of believing that this will work out, believing people will help.
00:38:29
Speaker
And if 20% of the people disappoint me, it's fine. Overall, I've got an 80% hit rate. And it's statistical, right? You keep sitting at the table for long enough, a hand will come that is good for you.
00:38:39
Speaker
Some hands will be bad, but eventually you'll keep building. So you have to stay in the game for long enough for the returns to come. It's just simple. It's just the attitude has to be, you know, make sure you don't waver too much and things work out. So you'll be actually surprised how much you can do without with little money if you have the patience.
00:38:57
Speaker
And this is, by the way, my second attempt. i ah We kind of tried it once before with ah another team, and another colleague of mine before. And we had our learnings from that and we move on. So you just have to stay home. Did you get any money from the government as part of India AI mission? Yeah, yeah, yeah. We have. We have.
00:39:15
Speaker
so ah yeah ah So this money from this entire allocation of resources from the government is a majority compute and then some cash, ah some so percentage cash. Like there's a GPU cluster which the government has set up or they are paying a private...
00:39:35
Speaker
GPU cloud. Yeah. Okay. okay Okay. Okay. What
NeuroDX's role in the India AI mission
00:39:40
Speaker
exactly is the India mission? I'm not very familiar with what is the goal or how much it is giving out. or Yeah. I mean, I can just comment on it from outside, ah but just being a little more familiar than any other person. Yeah.
00:39:55
Speaker
I think there was a realization, this is my my read of the situation. You remember when Sam Hartman came and just basically said it's hopeless for you to build. it Yeah, yeah yeah right right. And so I think people kind of were believing, yeah, I mean, ah it's all fine to be patriotic and be, you know, angry. eager How dare he say it? But at the end of it, the people in their heart of hearts knew he's right. right But when Deepsea came, things changed.
00:40:21
Speaker
When DeepSeek did something with very little rule resources, everybody then realizes that the cat is out of the bag. Now you have to do it. You can't afford to stay quiet because it is possible.
00:40:32
Speaker
right And so, although India AI has been around for longer than that, I believe, I think there has been this recognition by the government that you need ah a very sharp focus on this area because this is going to change everything we do and how we do it.
00:40:48
Speaker
ah and they have been doing a commendable job. I mean, I used to work in policy ah years ago. I was very serious about public policy and I spent two years doing it.
00:41:00
Speaker
I came from US and was very patriotic and you know the Anna Hazare time. yeah okay have worked yeah So I have worked with ah at the policy level and I have an idea of how systems used to work, right how a ministry would work and how things used to work.
00:41:16
Speaker
And when I now go again back into the system, Maiti and India, the kind of support and the kind of fervor and the intent and the zeal to do things, to get things done is incredible. I mean, there are certain things in the system which bring in inertia on their own. right It's such a huge system and there is bureaucracy.
00:41:37
Speaker
Things have to go through many people and you can't do much about that overnight. But if everybody involved has that focus, right, to, you know, we have to get this done, then it can move. And I have seen it.
00:41:48
Speaker
But ah what what exactly is India AI mission? It's it's essentially ah grants plus compute to AI champions. Yes, yes, yes. I think they are working at various levels. So even students are being given grants at a very small level, like PhD students, and you can write proposals and we'll fund your research at that level as well.
00:42:08
Speaker
All the way up till us, right? So we are at the topmost layer of what they support. and everything in the middle, right? And so what they're trying to do is make these GPUs available for everybody to use. So right now, academia, I think that's what they're hoping to do. I'm an academic, right? So I have an idea to build something out, which is of a grand scale, but I can't because I don't have the GPUs.
00:42:29
Speaker
So they are trying to make this pool of GPUs available at a very small amount, a very inexpensive amount, ah and have at least the academics, then let's say some industry people start building on top of. So they have taken on this challenge of making this very fundamental resource available. It's not like now like power. its You don't have power, you can't do anything.
00:42:50
Speaker
So I think that's what it's trying to do. And these 12 AI champions, ah of which you are one, do they get, like, say, business from the government? and It's not an as assured. the The relationship is not built built on that you'll get the business for sure. ah But, I mean...
00:43:10
Speaker
we ah We are not looking at it in terms of business for the government. We're looking at it that in a way that where we feel that we have to serve ah the government. We have to serve in a public capacity. So our models need to get into primary healthcare care centers.
00:43:24
Speaker
The conversation happens like that when we talk to people, right? That we are building from taxpayers' money. Although it's not grant, by the way. It's, ah we have to return part of the money. It's equity.
00:43:35
Speaker
Yeah, is there ah we haven't figured out exactly how it's going to be. Yeah. Yeah, yeah. So it's not fully like it's just been given away. But we still understand that this is something that has its taxpayers' money.
00:43:48
Speaker
There is a very keen awareness of the the mission and thus the kind of sensibilities with which this money was given. We also met the prime minister and ah um he invited us very graciously to his home, all the companies, and gave us two hours of his precious time.
00:44:05
Speaker
So ah I think it dawned on all of us that this is something that is beyond just pure profit making, at least on our team, it definitely. So ah we would want to red return to the government. So the first people we would like to serve is the public, right, in the primary health care center. So we are in discussion with various state governments.
00:44:25
Speaker
We'll hopefully try to get into central hospitals. And it's not like they're not thinking about how much money we'll be able to make. It's just like, let's just go ahead and do it. Okay. Okay. Your first product is the epilepsy detection. Yeah. Yeah. How are you selling? Are you selling a software? Are you selling a solution like a full stack solution with an EEG device and everything?
00:44:47
Speaker
Both. Both. So if, ah so ah what will happen is most likely ah the government ah will ask for a full stack solution from us. Like, And then we would be willing to provide that. We can talk to OEMs and get that done.
00:45:00
Speaker
But if somebody wants to use it today, and hopefully we'll be having the product commercially available very soon because it has to go through regulatory approvals, which is not is not enough that we have done our tests.
00:45:11
Speaker
right It has to be validated properly. But by product, you mean your software which will read an EEG and spit out a diagnosis? Yes. Yes. So that has to be a given regulatory approvals by the government. There is a body in India, CDC, which approves, ah which gives regulatory approvals on these products. And that then is ready for commercial. We have a testing license, so we are testing it.
00:45:35
Speaker
But when you have a license to be able to sell, At that point, ah anybody in any given day can just simply have an EEG, pull out underlying EEG from it and dump it into a model. It just takes five two, three clicks and you can pull it out. right um But we can also do a full stack solution. We are happy to do it. It won't take too much effort. So we haven't finalized how that deployment will happen, but I envisage that for the government, it will be most likely a full stack solution.
00:46:07
Speaker
Don't you feel like you are running a race with one hand tied behind your back with so little money? you know, now, for example, ah with money, you could have done these various tests of ADHD and depression and so many other use cases and then launched like a very, very compelling ah EEG in a box solution, whatever, which... We do feel like that.
00:46:34
Speaker
Yeah, we do feel like that. And we are actively raising our arm. So, ah yeah, so ah the the trouble was that ah for the longest time, it was, we talk about, I am being a little critical here, but there there are good reasons for it ah in terms of why situations, the situation is like this in India. We talk about deep tech, but deep tech is absolutely missing in India. It's not there. When I say deep tech, I say, are you building things that don't exist in the world, but you will be the you will be the first people to make it available to the world?
00:47:05
Speaker
That's In fact, India, and I'm i'm being very ah controversial here, in the last 100 years, or let's say even 200 years, I caused this as a challenge to my own students at IIT.
00:47:16
Speaker
Have we built anything that the rest of the world is using? So you get down at LAX, Los Angeles International Airport, find 100 people and ask them everything that they're using, even a single thing that they're using that fundamentally came out from India.
00:47:32
Speaker
I mean, if I look around this room, the paint, the light, the everything, the way it was, the metallurgy, everything is Western, right? We have made modifications on it. We have built on top of it. We have done great work. I mean, we have sent ah missions to Mars.
00:47:46
Speaker
But did we provide the fundamental tech to do things? Right. No. The answer, in my opinion, is no. The last time we exported tech to the West was yoga. Exactly, right? Exactly. So ah ah they this that's why i say last hundred years, even if we don't want to be aggressive.
00:48:05
Speaker
So and deep tech is that. For me, that is deep tech. other Everything else is tech or shallow tech. Let's not call it deep tech. Just because you're working on technology, you can't call it deep tech. Now, I understand India doesn't quite have that appetite. The financial infrastructure doesn't quite have the appetite because we don't have enough people who have shown how it can be done.
00:48:25
Speaker
So the longest time we people could see what we are doing, but they could not back us. It's just yeah the appetite isn't there. They just can't invest in a company which is a research company. We are a research, frontier we hope to be a frontier AI lab which wants to take on Neuralink, to be honest.
00:48:40
Speaker
which want to, Neuralic in the sense that we want to say, hey, what are you doing invasively? We can get, you know, 70 or 80% of that done non-invasively. And we can make it available to a person sitting in a village. Doesn't have to go and do surgery, he doesn't have the money to do surgery and all that.
00:48:55
Speaker
So if you are trying to build that company, you can't have kind of financial rounds that happen in India. It just simply is not feasible, right? How much are you looking to raise? We are looking to raise 10 million right now. Which is peanuts. and I mean, that won't move the needle for you. But but but no, we can still, it is peanuts. But you have you know you have to see it in the context of what we are trying to do. Even that, for a pre-revenue, 10 million in India, people have told us to our face that it's not possible.
00:49:22
Speaker
You have no earnings. Yeah, I can imagine a pre-revenue today in India, 10 million. But if you compare with AI labs, I mean, 10 million for an AI lab, just Meera Murati, OpenAI CTO, she's already pre-revenue, just purely on the idea, already worth a billion dollar plus. They're thinking machines or something, I think that company is called. Jan Lakoon got one a one billion valuation. He left, ah or you know, Meta. Yeah, exactly. Exactly. Purely on the idea. Purely on the idea. Credibility of the team. Right. right so
00:49:59
Speaker
And I was saying, yes, okay, in India, we don't have that credibility. All right. We have taken on the challenge to do that credibility. work We are in it for the long haul. And to be very honest, and this might sound a very very cheesy, but why with this company, it's not so much that I'm trying to solve a healthcare care problem. What I'm trying to get to do is show people after us that you can win a company like that.
00:50:20
Speaker
To show investors that India can, you can invest $10 million dollars in a company if they have all the things, you know, if you have green flags, which we we do, and learn to identify those green flags,
00:50:32
Speaker
what What are the green flags that you have? ah The green flags is ah first of all, the great amount of data we have, right? Nobody in the world has that kind of data. ah the We have already a team which has built the state-of-the-art model, ah ah foundation model on the output rate. No, but how do you prove that your model is good?
00:50:50
Speaker
Is there a mathematical? There are benchmarks. What kind of benchmarks? So there are, ah like, how do you do benchmarking in language models? You give it some challenge. The model is given some challenge to solve and see how well it is doing on that.
00:51:04
Speaker
Yeah, I remember there used to be like the chat GPT-3 could not clear the bar exam, but but by the time it was chat GPT-4, it would clear the bar exam. So it was as smart as a lawyer.
00:51:15
Speaker
ah So what is the similar benchmark for you, for Manas? So for us, it will be, let's say, some kind of a BCI task. some set of data which does a brain-computer interface task. Can you detect this, for example? Can you detect epilepsy, for example? with what is ah how How accurate are you in if you have a data set on which you can do epilepsy? And currently, are you are you more accurate than a human being who's reading the epilepsy data?
00:51:42
Speaker
ah The question, the immediate answer is yes. And ah I will nuance it by saying it's unfair to ask that question to the human being because the human being is not capable of doing what the Mushi is doing.
00:51:54
Speaker
So I would say that's an unfair comparison with the human being. I do believe like for a lot of x-ray kind of use cases now, AI is better at reading x-rays than humans. Yes, but the human can read an x-ray. So the human is trying to look at a broken bone. So is a machine.
00:52:09
Speaker
But in our case, that's not the case, right? Our our machine is looking for things that the human is not having doesn't even have the capacity to look at. ah okay the it is so microscopic the fluctuations in the graph that a human cannot read it this is unfair got it got it got it okay it's unfair to compare it with humans so yeah okay but but your epilepsy product what is the efficacy or accuracy or like is it like 99 accurate or 95 accurate what would like Yeah, it's it's so what happens is that we there are two versions. So you have to be very precise when you're talking about medical stuff, exactly what you're trying or claiming to do.
00:52:49
Speaker
So there is one thing that you can do is just take the EEG and just based on that, without anything else, can you tell whether the person has epilepsy or not? On top of that, you can add few very simple questions that any any MBBS, in fact, anybody can ask.
00:53:06
Speaker
Did you, for example, have a frothing after the event, whatever event it was? Did you have incontinence, for example, because some people lose their bowels or like ah urinate, right?
00:53:17
Speaker
Very simple questions. You can ask those questions and augment a reasoning model with a EEG model and then accuracy becomes even better. So the question has to be framed in the right way. So if you were just to take ah ah our EEG model standalone, we can hit 80 to 90%. Okay.
00:53:35
Speaker
okay And with the additional information of those questions? We get 95%. We get 95%. we get ninety five percent Okay, okay, okay, okay. ah Now, this is like what is already there, and I'm assuming now your goal is to...
Commercialization and global expansion goals
00:53:53
Speaker
I mean, you have two choices in front of you as I see it. One is ah commercialize this, so go all out in terms of getting this to be used at medical centers or primary care centers or whatever. Or the second is to...
00:54:08
Speaker
oh build Manas to be able to do more of the BCI work to, for example, judge emotions of people, et cetera, et cetera, and more of that cool tech. So which side are you more focused on? Are you more focused on cool tech or are you more focused on commercializing?
00:54:26
Speaker
um The difficult answer, it's it's very astute question. And this is a question that we internally obviously debate a lot. And ah so ah the the answer to that question is our primary offering is our model in terms of a company, what we are offering Manas, right?
00:54:47
Speaker
That is the central theme of our company that will be getting obviously a disproportionate amount of effort. Sandeep, our CTO, is supremely focused on building minus two, minus three, ah eventually you know being able to even ah show what you are imagining, possibly.
00:55:07
Speaker
Push that. Can you see what you are imagining, for example? So that is the central focus of our technical team. However, i personally am ah ah taking the charge myself to not lose sight of the financial aspects of it because I've been reminded of it so often. And we have to accept the realities of the place we are in.
00:55:30
Speaker
Nobody is going to hand us the kind of money that we want to be able to truly pursue this field the way ah we have just discussed. And obviously we have some responsibilities to the government which has given us this kind of money. I do feel, as I said, there is a great sense of being giving back to the to to the people who have given us this money.
00:55:49
Speaker
So I am personally, in my if when we think of our goals, I am personally more intent on operationalizing all the gains that we have done, getting this ah out into the market. Because I am also from a technical background, ah I am fully capable of, you know, switching back and forth in terms of bidding out the commercial sides of it, right? The clinical side of it.
00:56:11
Speaker
While I also handle all the executive roles. So, ah yeah, so... Overall, to answer a question, we are not losing sight of Manas. Manas is our main goal and we will keep pushing it as far as we can.
00:56:23
Speaker
But the epidepsy and all of those things, we hope to hit that brain scan in two years. Our internal goal is to have a full suite of ah tests available that we are ready to take abroad actually.
00:56:37
Speaker
Have done it in India and we are ready to go for the FDAs and Asia market with that test and ah let them know what you can take and do. So what you're saying is two years down the line, you would hope to be able to feed an EEG of a 60-year-old man into Manas or into whatever software and tell him that you might be getting early onset of whatever Parkinson's or whatever. like Early onset of, let's say, there is some cognitive decline beginning to happen. So cognitive decline can happen for other reasons other than dementia.
00:57:13
Speaker
So I would say here's the report. It suggests that there is things that you might need to kind of start taking care of. Go see a doctor. In the West, just getting a doctor's appointment with a neurologist takes months right now. What we take for granted here in India, right, I can go and see a doctor within two hours.
00:57:31
Speaker
That doesn't happen in the West, right? Even if you have the money. So what we are trying to offer is there is globally. So in India, what we are trying to do is fill a gap of service in terms of ah at the front lines. And in the West, what we are trying to do is fill a gap of service, even where there are doctors available, but just simply the access is not there because... you we can't get appointments.
00:57:53
Speaker
So a person can walk into our ah clinic and the it's like a blood test without a print break. Get a test done, same day report and have some sense of control over what is happening with them and then they can start, you know, moving to specialists grounded in more knowledge about their health than before.
00:58:12
Speaker
I want to come back to the funding discussion. Someone like a Sarvam, again, pre-revenue, managed to raise 40, 50 million dollars. Is it just that they were better salespeople, better at telling their story? Yeah, I mean, they were, they're falling at the, following at the heels of OpenAI, right? They're doing the chat GPT for India, like Indian languages.
00:58:31
Speaker
The understanding is there. This model is successful in the West. They are doing it for India. The risk factor is gone. they They don't have to prove the concept. They just have to prove execution ability. Yeah.
00:58:43
Speaker
The key is, can we be the open AI? When open AI was being founded, when Anthropic was being founded, what money they got, that is what we are at, right? They came up with something that nobody knew whether it didn't work out or not.
00:58:57
Speaker
Infosys funded them at that time, which was impressive. Like they were the first funders, the first list of funders is Infosys. Yeah, I remember reading about that. Yeah, yeah they had the farsight. So there was, ah I mean, Elon Musk and a handful of people who were there. They were individuals, right? And one of them was Infosys.
00:59:14
Speaker
So it takes ah belief and some sort of kind of courage and arrogance also at times to fund something like this, which the West has built entire civilization on.
00:59:27
Speaker
That is fundamental to their nature, taking those kind of risks. You're not quite there yet. But as I said, we are willing to bring that flavor here as well. Have you started pitching to VCs? Yes, yes, we are talking to a few VCs. And what are you hearing? um So we are in talks with, so they like the idea, obviously, but ah um what I'm seeing is people who have the understanding of, first of all,
00:59:54
Speaker
of the the full scope of how this is going to impact the whole world and maybe have access to the Western capital and market, they are more open for these discussions. ah Indian VCs ah is just outside of what they have ever done.
01:00:10
Speaker
they I think they get compelled by the idea. ah But after a few discussions, they ah they just can't feel secure enough in terms of backing this. that's my I mean, we have just started. We're just beginning on this journey.
01:00:25
Speaker
And to be honest, our our team is very clear. We will not scale down our ambition ah just to raise a few cheap dollars. I'm willing to die. But I will not scale down my ambition. I'll take it as far as as it needs to go. And maybe few people have to fail before other people succeed. I mean, things don't happen easily right? You have to build on top of ah giants, right? And so if our failure has to inspire the next round of great companies, I'm fine with that.
01:00:53
Speaker
But I'm very clear, I'm not going scale down my ambition just for raising a couple of million dollars. And I'm so thankful for the government that they saw it. I mean, imagine nobody else saw it. And among Sarvam, I'm one of the companies. I'm the smallest ah company in that cohort of 12 people. And for them to be able to recognize what is being done here, ah it was ah very, you know, we feel and so grateful for that, for them to be able to identify what we are trying to Amazing. Amazing. Thank you so much for your time, Siddhartha. I wish you all the best.
01:01:22
Speaker
Thank you. Thank you, Arshad.