Jake Aaron Villarreal: Welcome to our podcast, From the Ground Up, where we interview startup founders exploring their journeys, their success, challenges, and lessons learned. We hope you'd be inspired in discovering what it takes to build a thriving startup. I'm your host, Jake Aaron Villarreal, and excited to have with us today, Raul Popa, founder of TypingDNA, which went through Techstars and has raised $8.9 million in funding from Google and Techstars Ventures, to name a few. Raul, welcome to the show.
Raul Popa: Uh, thank you very much for having me. Sure. It's great to be here.
Jake Aaron Villarreal: Thank you for joining. I'll give just a little background, so TypingDNA was founded in 2016 in Bucharest, Romania, and more recently moved to New York. It's built an AI-driven technology that it says can recognize people based on the way they type, both on their laptops and mobile devices. Incredible technology, Raul. Uh, walk us through this. How, how did you even come up with that idea, typing biometrics, and what is that?
Raul Popa: Well, uh there were a bunch of uh, of other people researching the space, so I was not the first one doing this. Um you know, there's um, there patents that were written even 30 years ago that are not now public space basically. So, but I mean, but [there] are things that are um, using statistics and pretty much uh basic stuff to analyze typing, typing behavior basically. Uh the problem with typing biometrics before was that uh even if it was a thing... so people type uh uniquely, uh everybody has a unique typing style. Um the technology there was before, like AI came to power, uh was not really that uh powerful. So to, to be able to use this type of behavioral biometric.
So I learned about this type of thing when AI came around as well, so I was playing with AI and multiple other projects, pattern recognition pro- projects and so forth. And at some point I stumbled upon this idea and I thought um you know, the idea of looking how people type and using that as a way to recognize people is great, but you need to add you know the AI um component to it in order to make it really work. And that's basically what I did.
Jake Aaron Villarreal: That's great. For people that don't know you, you've had a storied career and are pretty well known [in the] AI space. Walk us through that a little bit. Kind of where did this all start from and where did you get the insights to really start trying to build products utilizing technologies that really were at the beginning of its stages in terms of what AI was, what it is today, and really how it's being used by a lot of companies. But you know, how it's being used by yours and, and really who's seeing the benefit from it? There was many, many questions there, I just... one.
Raul Popa: Um so I mean I, I believe uh, I'm not necessar- that uh, I don't know. I didn't do that much uh, that much uh in the AI space as a lot, lot of other people I know. I didn't contribute to the space that much, but I was playing with different uh machine learning and AI technologies uh you know more than 10 years ago, uh try to build products with it when almost nobody really believed it's a, it's a thing. And try to dive into different problems and figure out you know, what you know type of solution from here can be applied to this type of problem and how can you mix ideas around.
A lot of the times you see researchers only playing with one domain, like vertical domain, one type of data, one type of uh problem, try to solve that and they stumble upon a really good idea, but they never port it to another problem. And there's another group of scientists or a lot of the time they're just you know, you know academic uh scholars you know trying to, you know figure out the particular domain whether it is medicine, where is like all sorts of problems. But uh most of what AI does is pattern recognition. Uh almost you can argue that almost everything AI does is, is, is a type of pattern recognition. So and, and in a way problems are very similar and solutions are similar. So once, once you start understanding that, it's very easy to like figure out, "oh this solution over here could be applied to this problem over there" and so forth. And that's what uh for me in particular was uh was interesting, and I started playing with also uh ideas in machine learning.
And at the time I was managing product uh in another company. When I, when I started using this a lot, and for example at some point we figured out with a, with a small team that um the behavior [of] users on [a] website, on an actual platform to create um um banner advertising actually. With uh, with um... the behavior was predictable and we were able to understand from the behavior in the first minute or so whether that user might be um becoming a premium user or not, or if they might need a discount, or if they like really want to you know get to the end product or just playing around and things like that. So being able to understand that, we were able to double uh revenue or maximize revenue. I mean somewhere you know like that, some, somewhere like doubling in about 3 months, uh which was unheard of back then. And I started thinking that actually, you know, this ability of looking at you know behavior patterns uh can you know improve everything. So basically can be applied everywhere. So that, that was my you know main um interest from the very beginning.
Now I was involved in a lot of um products especially in the user experience side. And when I stumble upon this typing biometrics um idea, I found it very intriguing the fact that you don't really have to you know provide extra information. So whenever you're typing in, in an application um you know that application can watch the way um the way you type and use that to you know authenticate you, to reduce fraud, to not allow anyone else who pretends to be you to um you know log in and those kind of things. So for me that was uh you know, I realized there's really no other way to do it other than watching people you know, you know using their devices and, and, and, and doing whatever they, they do online. And it's really, you either become very intrusive like using [a] camera or microphone, or less intrusive like looking at um the keystroke. Especially the timing between the keys uh and how long the people press the keys. So for example, we're not looking at what exactly you're typing actually, we're, we're discarding that data uh in most cases for most of our algorithms who don't really use that at all.
Jake Aaron Villarreal: So if you're a company out there and you're a remote-first company, you got people all over the US, maybe international as well, how can your technology help them to make sure that the people that they think are working for them are working for them? And what does the technology do to alert them if it detects something that they believe isn't the right person?
Raul Popa: That's a great, that's a great question. So uh what, at TypingDNA, we developed a solution that does exactly you know solves exactly the problem you highlighted. It's called TypingDNA ActiveLock and what it does basically it runs on the computer itself, so it's not in the cloud. Runs on the computer itself and it's like a watchdog. It just learns how you, how you you know behave, how you use the computer. Uh we're looking at the way you type, we're also looking the way you use your mouse, but we're not sending data, data anywhere else. It just sits on the computer, is analyzed on the computer. And if it's somebody else, you know, all of a sudden you know at your computer trying to, to use it, um the application knows and then it alerts a, a central system.
In the same time every single interaction with that computer is authenticated. So it's not just like you're doing your job or you're sitting in front of a computer and looking at a video or playing around you know some, some game or whatever. Um the application knows exactly what you're doing and is able to authenticate every single action. And that is really crucial when you have people working remotely, if you're like a public company, if you're a you know, even if you're a small company but working with government or working with other public companies or other companies that for which you have to provide a certain level of um of service. So uh in all this situations you want to make sure that the employees um the employee computers are being protected all the time.
Um there's many use cases uh that uh um around this, this topic. So not just you don't just want to make sure that you know nobody like takes over that computer and starts using it for bad. Um you, you also don't want to allow you know people in, in, in your employee's uh household to use the computer. Because maybe on that computer there is sensitive data or important data that should not be seen even by you know your employee's spouse or children, or you don't want them to you know by accident delete something or send something to someone or click on a virus or just navigate to a website that could infect the computer and then maybe the entire organization. But there can be situations in which an employee would just hand the computer to somebody else to do their job. And actually this is a very prevalent uh uh uh situation, problem in um you know, in many parts of the world, but for example you know BPO companies uh that outsource uh services to, usually to like Southeast Asia you know, um East Europe and other parts of the world like South America and so forth. Where typically jobs are you know, you, you can pay less for same um type of uh job. Um these people usually work remotely and work from home and if they can, they cheat a lot of the time, they allow other people to do their job. And a lot of the time, you know, it happened to me so you're talking with somebody who does support from somewhere in Southeast Asia let's say, could be from anywhere, but you realize that person doesn't know well the product they're supporting and you know a lot of the time we arrive to the judgment that you know "they were not trained well" and all that and you know "support from India is bad." But that's not correct. Correct is that that person doing the support is not well trained, and a lot of times actually it's not the person that was hired to do the job actually. We know this from our clients.
Jake Aaron Villarreal: Yeah, that's a hard problem to solve, but if you have technology that can help you do that, I think that's an absolute victory. Yeah. Um what other areas... I know there's a lot of different use cases that your product is helpful. What other areas has it been used where you've got customers that are paying for it and they're seeing a lot of success out of it?
Raul Popa: Well there's many good things the technology can do. Uh you know one of the things where we have uh, we became like the de facto technology is student verification. So also you know remote, you have a lot of people who want to learn so they want to, to you know, you know do the entire learning process online and then also take exams online. And you know whenever you know, in the end you get a s- cert- certain certification, you want to be very strict about it or as strict as possible to make sure nobody is going to get that certification without putting in you know the hours to read and learn the material and also taking the exam while they're supervised by some sort of technology. And um, supervising an exam, it's called proctoring, that's the official name for that. So and then there's this big public uh, not public company actually, I think a couple of them are public companies, proctoring companies that work with thousands and thousands of universities and open platforms and so forth. And these proctoring companies, um some of them actually, a large number of proctoring companies are using either our technology, mostly our technology, but in some cases they develop their own or they use an alternative technology to verify students when they take exams from home.
And some, some cases you want to verify students even when they're learning to just make sure that they, they did their um, their um the learning, they read the document, they sit in front of a computer for five hours and, and, and looked at your video and every now and then you ask them a few questions to make sure they're awake and aware and, and learning. And uh we also have DMV schools using our technology for exactly the same reason. So um and there's many, many, many, many companies in electrical engineering only, that only real estate professionals only. So particular verticals in which professionals need to be trained and then there's companies that want to hire those professional so forth and certification needs to be valid. So you need a way to verify these people and a lot of the times uh proctoring companies use typing biometrics, especially TypingDNA for that. So that's uh a use case. Maybe a lot of uh people in your audience already seen um many, many, many years ago Coursera was I believe the first to introduce this type of verification, but then many other platforms uh started using it.
Jake Aaron Villarreal: What's the accuracy of this? So, say first and foremost, how do you get my DNA print, or not my DNA, my typing biometric print that it's me? It's Jake Villarreal that's typing. Do I have to go through a process so that you can verify first it's me and then apply that, and how accurate is it?
Raul Popa: I mean it's uh, very good, very good question. So it's uh the technology um it's very easy to, to test. First of all you can go to uh on our website typingdna.com and there is a bunch of demos you can try for you, you know to play with it and, and see how it works. And um you know you get, you get a glimpse of um how it's been done. It's very simple. Basically every time you press a key and then you release it and you press another key, sometimes you press the next key before you release the first one and so forth. Um those timings between the keys and how long you press each key, they make up a typing pattern. And one could argue that you know uh they type differently today than they did yesterday and all that. But in reality it's not true. It is differently, but it's just slightly different. And our AI was able to learn...
So statistics can learn all the way up to like let's say 80, 85% accuracy, which is, is not enough. You need AI to, to reach you know below um, beyond 95%, 99%, 99.99%. Depends on how much data you have prior to that interaction. So it all depends how much data you have from that person you know from prior interactions and um how good an algorithm you're using. So the AI-based algorithms that we're uh that we developed uh can reach all the way to 99.99 something like that. Uh really depends also on, on the amount of text you're typing. If you type only one letter, you still typed something but it's not enough. Or only that letter 30 times, it's not, it's 30 characters but it's the same. So um it depends on these kind of things. And generally speaking, AI depends on, on many things like that. Like for photos depends on lighting or you know how much of the face is, is, is uh available, visible, how much is covered and so forth. So you get all those um all these uh questions.
"How accurate is the technology?" We, I can give you a number but it's going to be in a particular box of uh conditions. So if these conditions are met, then we have this uh particular accuracy. For example when you have username password or email password, um about 30 characters in total, email plus password, then at that point we are uh pretty confident we can reach over 95%, which just uh two previous enrollments. So two times you log in, third time you log in you already have over 95% accuracy. The idea is, but that's with very small, small amount of data. For example the watchdog that I told you that sits on your computer, that is more than 99.99. That actually is uh well, almost always um so far we don't have a um a opposite uh situation where we never uh were, never able to flag a fraudulent user. We almost always flagged a fraudulent user. Uh the trick there is sometimes you can very, very quickly know it's a, an imposter using that computer. Sometimes takes a little longer, but every single time so far with everyone we test this technology, we were able to flag the right person um to be, to be the imposter and that's very, very hard to do.
Now going back a little bit at the accuracy um when we talk about like face recognition, fingerprint recognition and, and you know, iris, things like that, we talk about one in a million or one in millions uh people would be similar to you. Maybe even one in a billion or something like that when you're talking about iris. Um the, the thing with typing biometrics, and I don't want to you know say otherwise because it's not otherwise, is that about one in 10,000 people type like you. So one in 10,000. That's not really that uh accurate if you think of it. So even with uh complex technology, okay let's say in the future we will be able to look at like sensory data more complex from a keyboard, like pressure and uh you know things like how long it takes to actually reach the bottom of the uh the key to reach the bottom when there... the first basically sensory um you know call this... the first time when we know that, that the key was even touched is when it reached the bottom. So let's say you know also when the key was touched and all that. So let's say you add more um information, then in that case probably is going to be like one in 50,000, one in 100,000, but still it's not you know unique, unique or 100% unique.
A pin on your card, it's one in 10,000. So they're 10,000 combinations, right? All the way from 0 to 9 for, for uh numbers, so there's about 10,000 combinations. The idea is every one in 10,000 people has the same uh pin that you have in your card. That doesn't mean that if somebody has your pin doesn't mean they can know who you are. There's no way to you know reduce one pin to a person. It would reduce a million people, a lot of people. And the idea is this type of biometric has its merit because if, if it, if it cannot be reversed to a person, to track a person, then it only proves what it needs to do. Proves ownership of who owns this credit card or debit card because they know their pin number. So it's perfect to use as a 2FA. It's perfect to use as a watchdog. It's perfect to use, and you know um online shopping experience, uh online authentication experience, because it doesn't say who you are. You can sell crypto and buy crypto with it and be on a you know dark web and stuff like that, and you can still add the security of typing biometrics. That is the cool thing, the cool part of it. Where face recognition, fingerprint recognition, iris recognition, voice recognition is uh more accurate and has the problem of being to privacy um too, too big of a problem from a privacy perspective I would say.
Jake Aaron Villarreal: I want to go back to one thing you mentioned about the fingerprint, actually the how far maybe you press down on the key. What if it's on an iPad or a mobile device? I mean, most companies and most people today they work on computers, but there also is a fair amount that you hop on your iPad, you hop on your mobile device, you respond to things. How accurate is it there?
Raul Popa: Um yeah, let me stand a little bit. Um so um on the mobile devices, the technology that we have is not exactly on par with what we have on desktop. Um the reason being is that you have a different type of keyboard, different form factor and um it's just touch keyboard. Uh so you type differently on that, on that particular type of keyboard. And actually there's six positions, main positions in which we know that people type. It's not just one, one type of uh typing on, on a mobile device. And I can, it's very simp-, I can explain. So it's like, you can type with one finger like that, or you can type one finger like that, it's completely difference. Or this finger, or, or two fingers like that. And so there's many ways in which you can type on a, on a phone. And uh, and because of that, the way you type in one, one style is going to be different than the way you type in another style is. And we're not, we were not able to create a, um an algorithm that will be able to authenticate the way you type on computer versus the way you type on a mobile, or between this uh six styles. Um actually there's three main styles between which we cannot do anything. The six styles have some similarities and we're able to something, but in the same way, if you record this fingerprint on a, on a touch sensor, touch a sensor, right? And then you want to authenticate with this other fingerprint from another hand or, or different uh finger, it will not work. It's the same way. Or iris. You record this eye and then you want to authenticate with this other, it's, it's, it's you but it's a different eye and different iris and different fingers, so forth. So that's the way it needs to be treated.
But uh now going for... uh usually people will uh, will type in the same way on, on a mobile phone. If you typically type in with one finger, you're going to typically going to input uh text with that finger uh next time. And especially when you like log in in, in an application like a bank application or let's say you're on Amazon, you're buying something and you asked to type a certain you know four words or three words or something like that, certain text or your credit card information, is, you're typically going to introduce that in the same way even if that you know, in a WhatsApp application you would type in a different you know position. So that's enough for us to do authentication very seamlessly without people realizing that um you know there's so many options in which you can actually you know hold your phone and type. That's one thing.
And another thing is that uh on mobile phones we also look at the sensory data from the phone, at the telemetry data. Accelerometer and gyroscope and basically all the forces you apply on a phone while you're typing. Um every, every movement, every small tilt and micro-tilts as we, as we call them, all those counts. So when you type in a key and you release, that you know also sor- of forces and rotation forces and you know translation forces are applied on the phone and they're captured so nicely by the sensors. You basically see you know, I think it's nine uh axes, so nine, nine dimensions on which like everything vibrates like that, right. When you look at all the sensory data and then when you put that, that information in AI, you're going to have even more accuracy or better accuracy that you have on a, on a regular uh computer. Um even if you just type with one finger. It's crazy how um how far you can go. Typing biometric on mobile device is a thing uh but it's a different thing than on desktop. And we have technology for both.
Jake Aaron Villarreal: You know, you've got a really storied career. I know that uh in our previous conversations you've talked about speaking at the European Parliament and um if, correct me if I'm wrong, was it specific to AI and kind of where AI is going? Um walk me through what was there that they brought you in to talk about? Because I think for the listeners, there's so many companies that are in AI and they're learning what it does and how it can be applied to businesses. But there's really experts too that are you know shaping policy and talking about how it's impacting you know different parts of um technology, but also person- people around the world. How, how are you positioned in the AI space?
Raul Popa: Yeah. So I think one, one thing led to another. So I started playing with AI in multiple ways. Then I went uh with TypingDNA actually went at the very beginning, I did some research, went to a business accelerator, met a lot more people and started to you know, you know talk at, at all sorts of events. I, I spoke, I lectured at different events in universities and um AI events like um, Applied, Learn Applied Machine Learning Days at um, at EPFL in Lausanne uh, or I talked at TEDx about AI as well. I, I talked in many other situations like that. I uh got some awards to be like number 10 or whatever influencer in, I'm from Romania originally, it's European, so I got some number, I don't know what number it was, in the top of like most uh influencers in artificial intelligence and all that. So ev- eventually people started asking me to come to different events and talk about what I know. And I started meeting other people. I, I started a group called the Romanian Artificial Intelligence Society and I now, I know pretty much everybody who is uh, who's working in the space let's say.
And so it was, was not necessarily that I'm better at this than a lot of other researchers that I actually know, and some of them are really, really good, um I would say better than me, because they only focus on AI. Um I focused on many other things and AI is just one of them. I'm very fascinated about AI, especially AI applied to cybersecurity and user experience. But uh but yeah, so um in, in the European Parliament there was one initiative from Renew Europe. Renew Europe is, I don't know if it's the largest party, but one of the largest parties, and um, to be honest it's very interesting because in, in United States you don't have um... you... I live in United States now, I'm a permanent resident here, so say, I should say "we." But uh, so in United States you only have left and right. You don't have the concept of the center, right. It's, and if, if you for example, you're a you know Democrat and you vote blue, and then you have a friend that voted red, all of a sudden you cannot be friends anymore. It's like, there's no way people that vote different can be friends. Or you cannot agree that maybe, I say "I don't like Trump, but let's say maybe one or two percent or something that he did was okay." Cause maybe I'm saying uh that is uh, that is the way you think in, or, or the thinking is happening in US. Everybody's either left or, or right. Nobody's center.
Now in Europe you have all these countries, like Germany has it, France has it. Europe Union in general has it. Has the middle party. And Renew Europe is what actually, probably if, if I'm not mistaken, is the largest party. And it's a middle party. It's a, it's what is called a swing party, because basically they swing either left or right depending on what uh they agree with. And actually they make the, the, the majority of the, when, when, when they vote something that law goes into action. So they uh started this movement about the AI, the future of AI, what we do with AI in Europe. And they put together a, a nice um event about future of AI, and I was invited to be a panelist on that.
Jake Aaron Villarreal: That's great. Yeah, um I know there's a lot of areas around discussion and regulation around AI, and I guess it'll figure itself out as it goes. I want to talk a little bit about you. Um you were in Romania, you got into technology, fast forward, you're in the US now. You went through Techstars uh, you got funded by Google and Techstars Ventures. You got to have something pretty good for them to, to back you. What was the biggest benefit you had going through the accelerator Techstars, Techstars?
Raul Popa: Uh well, Techstars is a great um, it's a great experience for every founder. I know a lot of your uh audience is founders. So let's maybe talk a little bit about this. Um so a lot of people think that uh there is a, only Y Combinator to start from there. Uh Y Combinator came up with this theory or this uh idea that good startups always come from Silicon... or they're built in Silicon Valley. So that's why Y Combinator is in Silicon Valley, is in San Francisco. There's no reason for them to exist anywhere else in the world, because good startups, really good startups only come from there. Which used to be true, if you think of it. I'm not saying it's, it's 100% wrong uh, or, or anything like that. And you know, if you can get in YC, by all means, YC is, is probably the best accelerator if, if you take just one chapter alone. So okay. But if you think of it, you know, Techstars has a completely different opinion. From where it starts, everything else is very similar, but they start with the idea a good startup can come up from anywhere, can come from anywhere. So you don't have to be in Silicon Valley to create a good you know generative AI startup, or cybersecurity startup, or, or anything like that. As you know, since you've been born in some other place, you're probably going to you know, go to school in some other place, you're going to, you, you know, mingle with other people and create a startup in other place. Why not, you know, um why not be uh being able to succeed there?
So the idea of uh you know that, that resources like are so scarce that you have to go physically in Silicon Valley so that you can you know start a successful startup, to dissipate. And Techstars realized this. Although Techstars and YC started kind of in the same uh period, I would say the YC thesis was more correct at the beginning, and now the Techstars thesis is more correct, you know for the time being, and probably for the future. So Techstars says "look, you can start a business anywhere." However, there's still some Techstars chapters that are more you know successful than others in fundraising afterwards after you go through Techstars, in you know type of mentors they have, things like that. But also the, I believe that's also the localization problem that you know, has YC. So basically, there's more funding in New York, let's say, than I don't know, North Carolina, you whatever. Um and um, which is fair to, to say. So we went through this uh Techstars New York chapter, New York City chapter basically, um which is I believe uh the most successful one in, in entire Techstars. There are more than 30 programs, but the one that bring[s] the most uh return for Techstars as, as far as I know. The other two that are very close are Boston and Boulder. Boulder is the one where in Colorado where Techstars is based from, bas- basically they're based there. And Boston is actually where they had the first accelerator. So they have all of, these three have uh more than 10 years of history of batches, um two, uh two batches a year, 10, 12 companies. So hundreds of companies in total in uh, in each of these three accelerators. Now I heard really good stuff about Techstars Berlin, Techstars London and so forth, so I would not say anything bad about this. But we went through Techstars um in New York City and it was [a] really good experience. I recommend it to everyone.
Jake Aaron Villarreal: Is it the network or the funding or the sounding board? What, what, what for you, do you, if there's one thing you can say that you left with going, "you know, this is really what they did for us," what would that be?
Raul Popa: Wow, so first of all it's incredible people that are working in, in Techstars, even now. So now it's, for example, current Techstars has a different managing director than the one that was uh when I was in, in uh in the program. Actually the person who's now managing director used to be um program director or something. The person next to the managing director I guess, not 100% sure, but uh was one of the main people there at that time. Now he's, he's, he's leading. Before that uh there were really, really good people that I met uh. Eventually I met with other previous or, or, or, or future uh managing dir- directors at, at some of these um, um chapters. Um they're really amazing people. People that could build or already built two, three startups uh, and then decided to do this. They're really good at teaching, they're really good at um...
For example, Alex Iskold, the, the guy who was managing the, the New York uh chapter when I was there, um it was a, he's a guy that very, very fast understands people and understands uh whose talent is what. And was able to very quickly look at different teams and say, "you're good at that, do that. You're good at that, do that." They were like confused, they didn't know what to do, and all of a sudden everything was clear and they started executing. Uh so sometimes you only need that. You don't necessarily need only the money and the badge that you're Techstars and all that. So, and it says, it says it's harder to get into Techstars than to get in Harvard, so it, in a way it's like an MBA if you want. The same with YC. YC is very hard to get into. So look, I'm not speaking against YC. I have a lot of friends that went to YC and I know it's a good uh, it's a good experience.
Jake Aaron Villarreal: When you started uh in Techstars and you had your thesis, how much has changed from your idea about what you were going to build to where you're at today?
Raul Popa: Well I think a lot changed. Uh I liked the idea of uh of applying uh AI to many problems. I found this uh typing biometrics thing that could reduce um the problems in user experience across different types of products. So not just uh... authentication is everywhere, and, and fraud is everywhere, so if you want to reduce that without additional friction, this is the way to do it. So I really liked that, that part of [the] thing. Um and um you know, when I started I was like very sloppy in what uh I was thinking that I need to do. But this, not, this is not the first business I was involved in. So um some of the things I got figured before, but not everything. So we needed to pivot, if you want a little bit, and create new technologies and other technologies. So basically when we named the company TypingDNA, we thought, you know, it's easy to name it like in a, like Google, right? So, or, or like Apple, or give it a name like Amazon that could be so big that like, bigger than anything else in the world, and at the same time, like or the X app for everything, right? And then you start small, and that's, that's a good strategy for some.
But we went the other way around. We said, "let's um you know, name, name it really narrow, like TypingDNA. So this, we're not going to do I don't know, face recognition or something as a primary thing, or we're not going to do you know um I don't know some generative AI that will build you know images whatever. Uh we are going to you know sit on this problem and figure it out. And uh you know figure out you know where this pro- this um you know typing biometric thing could solve different problems and uh, and not you know diverge too, too much." And I guess this builds a sort of resilience into how much uh you're going to focus on a problem if you from the very beginning narrow your um, your scope. So the name helped us narrow our scope and be very focused on what we do. And I guess at this point we're um, I think you know is the only thing that I would brag a bit is that I believe we're the best in, in what we're doing, by far, in typing biometrics.
And we not only did for um you know authentication, we also try to see what else we can do with this type of biometric. One of the things we're able to do is tracking mood. So we have an application which is actually free, any, anyone can install it on their computer and watch their moods. It's like a fitness tracking, fitness tracker for uh for your computer. You install on your computer and then every now and then you look at how your moods um you know progressed, and how happy you are, how tired, how stressed, things like that, how energized. And you know, think of it like you, if you in front of, sit in front of a computer for like, or stand in front of a computer for like um eight hours a day or a few hours a day, uh in that time probably your, your typical fitness tracker will not record too much you know um relevant data. But you're going to type a lot. You're going to you know be in different um situations when you need to type, whether on Slack or email you know documents and so forth. So the way you type on that computer every, every single hour will show how um, how your moods progressed. And you can go back and realize, "oh I actually enjoy doing this more than I, than I enjoy doing that," or "I, I typically am more energized in the morning than in the afternoon," and so forth.
Um the mood changes very fast. And, and many times a day. A lot of times you, you, you, you, you know, a regular person would think that um you know "this day I was stressed, this day I was happy." This [is] not true. As the f- the fact is, and as we watched it uh uh fluctuating, uh you can be super stressed in the day and then next hour your, your you know stress is completely gone. So it really depends what you do to you know relieve your stress, the tension, you know get you know energy going up and all that. So and you can basically you know try different things. Diet, you can try you know taking a nap, drinking a coffee, going you know 10 minutes for a walk, standing, sitting. You a me. And uh, and you see how everything improves. And uh if you're a person working in front of [a] computer I, I believe it's a great thing to do. So typingdna.com/focus I think it's the, it's the URL let me, let me try see. But I guess that's the thing.
Jake Aaron Villarreal: I think that's fascinating. As you pull that up, you can actually track your mood and your energy and also the patterns of how you operate. If you could do that through typing, I think that's amazing.
Raul Popa: We did that uh because we thought it's possible and I wanted to see, see how far we can go. But also we wanted to collect a lot of data. Not through the app, we have a public uh research app, so separate than the app itself. The app itself will not send us data from you, uh but we wanted to collect a lot of data for continuous authentication because we, we wanted to build this watchdog that I've told you about that has to learn how you type for a longer period. So we said, "look, you want to participate in our public research? We're going to use the data for continuous authentication, for mood tracking predictions and so forth. And we're going to give you more uh moods to look at." So by default it's for free, you, you see uh three moods. But if you, if you participate in the public research which is in a website, in a webpage, so it's not installed on your computer itself, the, the research. Then you get three more moods. Um...
Jake Aaron Villarreal: What's the, what's the page? Where would they go to find [it]?
Raul Popa: typingdna.com/focus.
Jake Aaron Villarreal: Got it, okay great. Really cool. When you started the company and where you're at today, um how, how, how big are you now and where are you located?
Raul Popa: Yeah, so we're not a really big company. We, we're an R&D lab basically with some marketing, some sales attached. So our biggest team is still in Romania. We have an office in Oradea and uh some people scattered the, the rest of the country, in Bucharest mostly. And uh we have some people full-time or part-time also in other parts of the world, including here where I'm, uh where I'm, when the company is based in New York City.
Jake Aaron Villarreal: What's the biggest challenge that you are going through, or you went through when you started the company, that you didn't know about that you've had to solve as you continue to grow the company?
Raul Popa: So um I think if you, if you're not making the company you know highly successful very early on, uh it's uh, it's very hard to raise additional capital unless you're working on a very novel problem that everybody believes is going to be huge, like generative AI today. So today if you're working on generative AI, then you get easy funding. Um you know, maybe a couple days, a couple years ago, if you were working on NFTs and web3, you probably get that kind of you know, free round if you want, easy funding without you know having to show like a lot of traction. Um I would say that in uh cybersecurity, um it used to be a sort of free round for a few years because uh there's no way to build cybersecurity uh to have uh traction before you build cybersecurity products. First you have to build them, and then, and you have to be really good at that. And then you, you basically get, get uh to sell it, and that's really very hard to sell because you have to build the trust.
Now, you don't have to just build the technology uh as in other domains. You just build, you just make a lemon stand type of startup. We, we make a lemon, then lemon, and you make sure you're in, in a path where a lot of people pass, right. And then you sell it for you know how much people want to pay for it. And that's what 99% of the startups are doing. They, they making a lemon stand kind of start-. You can argue with me, but that's what they're doing. And then there's startups that uh have to, have to do something else. Like in our case, in cybersecurity startups in general, you have to build trust first. Because a large company will not agree, it's like agreeing to work with an imposter, with a, a thief or something like that to protect you from thieves. There's no way, there's no way. So if you're, if you're going to use a cybersecurity solution, you want to use it from a very respectable vendor that is very good at what they're doing, have all the certifications, can you, you can try that technology and all that. So now if you have also very innovative technology that only appeals to, appeals to early adapters or innovators um, like in our case for example, AI cybersecurity solution based on the way you type. Then a very narrow, you know, very small amount of um potential customers you're actually targeting.
So you need to have the right investors that really believe in that and are able to put more money before you actually start to see a lot of revenue. And this is... um I'm explaining this because I think this is valuable for many other types of startups, not only cybersecurity. But there's a few types of startups that are not typical lemon stand startups. Uh and I, I'm not saying that in, in any uh bad way, I think it's a great thing if you, if you found your lemon stand type of startup and make, can make a lot of money without you know amazing technology and amazing challenges, that's fine. But uh for, for startups especially those in like cybersecurity, um this is one of the biggest challenges. How do you get the right investors to believe in you before customers believe in you? How do you build trust so the customers start believing in you? And how do you start um you know, generating um revenue? That's very... it's a few challenges, very, very hard you know, that are on top of how hard it is to actually build the technology. So a lot of people think that the hardest thing is to build technology. Yeah, I, I beg to differ.
Jake Aaron Villarreal: So you think that it's building the trust with your technology is maybe harder than actually creating the technology. Because you got, you have to have clients and customers, maybe even the government, to buy into what you're creating that will be beneficial. And so from that capacity, I can understand that trust is maybe as important as anything, especially in cybersecurity for sure.
Raul Popa: Yeah, that's, that's what I think. But I think also in generative AI. So for example, uh ChatGPT was around for a year or so until it became public because they were afraid that people will not like it. Not that will not like how smart it is, but uh will be afraid of it. Will not want to use it. And will be, you know, seeing you know the capability. So it was pretty much a trust issue, but a different type of trust issue that they were facing. And you, you know, all of a sudden when they uh they decide to go you know out and tell everybody that they have this kind of technology, I think it was a good, a good moment. And then Microsoft stepping in you know, buying shares in uh in OpenAI, they basically like uh infused the whole generative AI and AI domain. Everybody's now interested to, to build AI tools, like, we're seeing with clients. Like, clients are more interesting to use our technology because it uses AI before... before that, trust me, this is, this is the truth... before this, they were afraid of using us because it used AI. So AI was not a factor that you would bring in a discussion when people will ask you like, "And how does your technology work?" "Well, we look at how people type," and all that. We were not talking about AI that much because they were afraid of AI. Now when we say we're using AI, it's the same thing that we, we did like few years ago. Just that now people trust it. "Oh, AI, the thing that you know, we already know it's good uh or is better than humans in many things." Which yeah, we knew that. Just that um you know, the, the, the common knowledge and trust in, in this type of technology was not there yet. And I believe now it is and uh, and it's a great thing that it is. Um although I have my fears about AI. You should...
Jake Aaron Villarreal: I have a question for you. You shared that you know when AI can do something that's superhuman, that there is some real value behind it. From your perspective, what does superhuman mean and where do you think AI is today? And, and how far do you think it could really go in terms of changing how we operate, the jobs we have, the roles that might be lost because of this technology and the innovation around it? Um, questions anyone... um yeah.
Raul Popa: I'm not sure if I'm answering all these questions because basically you're putting this open question out there, so I just pick on whatever I, I want to answer, alright? Whatever I think is valuable for your audience. Um...
Jake Aaron Villarreal: What is valuable for your audience, I mean I think understanding, you know, there's a lot of companies that are building in AI. Are they building the right type of solutions? Are they you know, where, where can it go? I think it's not just about entrepreneurs, it's also about engineers that are out there listening. You know, what, what, where do they want to be building? Where, where's the real opportunities for them?
Raul Popa: Yeah, well um yeah, so I believe in what you started the last question with um. You was, you was, you were uh you know talking about the idea of superhuman AI. Um it used to be that people, especially like top scholars in the AI space, they thought that um, will do very well in the non-s- superhuman space. So basically this is [a] famous quote from Andrew Ng, who's one of the, the top teachers in AI, has a course that was done by a million people or so in machine learning at Stanford. So, and uh he says that uh AI will be very good at replacing people at tasks that it takes less than 5 seconds for us to do. Everything else should, we should not even approach. If it's a, if we're thinking that AI should be used to, to do something that a person will take more than five seconds to do, then it's too much.
So then in those 5 seconds kind of task you have self-driving. If you think of it like um each turn separately or each decision separately can be done under 5 seconds, or that would be one of the toughest problems. Or diagnosing something based on looking at some images or you know um you know MRI images or, or any, any type of images. So those kind of things fall under the, even if it's a professional it takes like less than 5 seconds to, to do the decision and things like that. Like face recognition for example, and, and there are many, many, many others, you can come with so many, so many domains in which AI could do something. But uh interestingly, um the ChatGPT kind of application, it, it's, it's, it's above that 5 seconds. So uh and we're talking about a prediction that was made by one of the top scientists in AI like five years ago, and now it's gone. That's uh quite crazy.
Um and we call this type of AIs that take less than 5 seconds let's say "non-superhuman AI," they're "human level AI." Not even uh top human level AI, let's say top human level AI would be conversational AI which takes more time. You need to think, you need to be aware of what happened before, you need... it's not just one decision, it's a continued conversation kind of thing, you know, putting together multiple things, making sense of the world and all that. So that's already beyond the basic human uh AI, it's the top human AI let's say. Um maybe only like some people would be able to do that and so forth, not everyone. And then there's superhuman AI. And about superhuman AI the, the things uh the ideas were, were uh were different. Because there were some uh domains in which without AI you cannot do anything. Like typing biometrics for example. I will give you this example because this is the easiest that comes to me, but there's many, many other domains. Like uh DNA uh um or u- different types of researches in DNA. DNA folding, all sorts of uh researchers, uh researches around this, this topic. Or medicine and so forth. Where even if you have uh a thousand hours and a thousand top scientists, you will not be able to solve those kind of problems. So you need a large computation capability and also AI to solve those kind of problems. Those are superhuman AIs. And there were uh there were um... this type existed also five years ago.
So when I said earlier that Andrew Ng said only the, only the, the basic human uh tasks should be automated with AI, he was not referring [to] those superhuman, only uh not the, the more you know top level human. So for example like management, management was a thing that was not considered easy to be done by an AI. Nowadays it becomes more and more logical that management is going to go uh very, going to be replaced very fast. Or even like I think it's easier to create a president AI than a employee AI. You know what I mean? So it's um, that's, that's what's changing. But superhuman AI was, was here before and it's here to stay and it's an entirely different domain.
So for example if you look at two typing patterns, you will not be able to say as a human uh who's who, or is... if it's the same person uh who's typing. Or fingerprints too, it's the same. You look at fingerprints, you can, you will not be able to distinguish and say "oh, this is the same person." You have to look so closely to all the details and um, to be able to compare fingerprints. So um uh and, and, and you have to look at so many, how do you, how do you recognize between so many fingerprints and all that? So uh at this, this particular type of tasks uh, we call them superhuman tasks, um AI was very good and I think it's becoming better and better. And um this is the type of way in which I think we'll see a lot of progress. But then there's also the s- the I, I, I kept calling it in this, in this podcast a "top human uh AI" because it, it require like for example um going through a document and summarizing that. It's not a task that any human can do. It's a task that some humans can do. You have to be very proficient in that particular thing to be able to understand, to be able to summarize well, to speak good um you know the good language that you're going to summarize in and so forth. And um that's not easy. So maybe like 10% of people would be able to do, like the top professionals, right? Um those kind of things I think uh is going to, is going to happen.
So a lot of AI you know doing things that you usually were done by professionals... professionalist, whether it is like summarizing something or being able to come up with uh the right I don't know, to look at an um contract or an agreement or some sort and, and figure out what's wrong, what's good, what, what needs to be there um in a contract without employing a, a lawyer for example or anything like that. These kind of AIs I think will be uh is a thing that um people should pursue more. Um for two reasons: first of all, nobody was doing this or very few people were doing this like a few years ago because it was thought impossible. And now it's, it's believed to be not only possible, but uh very possible to, to you know, the degree of accuracy is considered even better than humans in most of these tasks. So that's definitely where I would put my money. And then uh obviously the, the basic human tasks, those are still there to automate with AI. And then there's um, there's super human AI where uh there's a lot of research being done. Which yeah, there's all these three domains. Maybe there are other domains, depends on how you classify them, um that worth uh you know looking at. But I believe the one that nobody paid attention [to] until like a year ago or so is this um top level professional human uh sort of level AI.
Jake Aaron Villarreal: It's fascinating, Raul. I mean the experience you've got and the knowledge in the space I think is amazing. Um before we wrap up here, is there anything that we haven't asked you that you want to share?
Raul Popa: I don't know, I, I believe no. Um, nothing comes to mind. Okay.
Jake Aaron Villarreal: I got it. I know we covered a lot here. Um so if, if people want to reach you... again, just share your website and where's the best way to connect.
Raul Popa: Interested in technology that, that we work with and we gener- we created and, and, and all that is uh typingdna.com. If they're interested to or commercially and all that, there's uh some buttons they can click there contacting us and learning more and downloading um information and playing with the demos and all that. And the Focus application that I, that I said is free for anyone to, to use it. Um not just to try, to, to use it on Mac and PCs. Um and second, if they want to contact me personally, they can do it on LinkedIn. They just look, have to look uh me up. My name is Raul Popa, and um they can also look me up by my company TypingDNA.
Jake Aaron Villarreal: Great, Raul. Thanks so much for joining us today and thanks for all the listeners that are listening. It means the world to me and, and to us. And uh again uh, as we continue to explore this journey around technology and startups and entrepreneurs, fascinated to see where AI takes everything. Uh this is Jake Villarreal signing off for now, but can't wait again to connect with you soon. Take care.
Before you wrap up, I want to give a big shout out to all the entrepreneurs that have joined to make this podcast possible, and for all the listeners for listening. It means the world to me that you chose to spend your time with us today. I'm your host Jake Aaron Villarreal signing off for now, but can't wait to connect with you all soon on the next episode. Take care.
This show is sponsored by Match Relevant, a company that helps venture-backed startups find the best people in the market. And they do it in three simple steps. First, they sit down with founders to understand their story. Second, they tell their story into multiple candidate channels. And third, they schedule interviews within 48 hours. Find us at matchrelevant.com to learn more about how we do it.
.jpg)
.webp)
.webp)
.webp)
.webp)
