LaunchPod - Rajiv Jain (audio only) === [00:00:00] Rajiv, welcome to the show today, man. Glad to have you on. Thank you so much for having me. I've been looking forward to this one because I think there's some really interesting things we're gonna dive into here. I've heard a lot of discussion around how good are models gonna get and have we plateaued, and I feel like every time someone brings that up, Anthropic or OpenAI just drops their next nuclear option, and it's insane. But still, I hear that, so interesting conversation to have there about really the realities as well as how you can next level some of this stuff with also bringing in really, really proprietary data sets as well. But the other piece I'm really interested to talk about is your background helping to deliver a lot of the enterprise-grade early assistance from Alexa and things like that, that were really precursors to modern agents, and how those learnings kind of parallel. But before we get into all that exciting stuff, how did you get to do such a collective amount of [00:01:00] really cool things over your career? Just quick background. I am SVP of Global Product Strategy here at AliveCor. I'm leading the product and the AI strategy, and I am deciding on what gets built, how we take it to market, how we make money. I started in finance- Mm-hmm ... and I've had, like, a nontraditional product management pipe, really starting in finance and then over a per- at Microsoft, and over a period of time seeing the industry while getting into strategy. And then from there, I founded my own company. And I've been involved with AI since 2015, really thinking about how you take teeny-tiny models that you can fit on custom chips all the way to Alexa at Amazon, where I built the enterprise ecosystem around Alexa, and then later on to AliveCor. We are now involved with the clinical algorithms that we deploy on custom devices that we sell to millions of customers globally. Nice. We always run into people on the show who say, "Oh, I had a weird path into [00:02:00] product," and I think those are often the best stories because you bring that wider context. So yeah, I mean, I'm really, really stoked to see how that path contributed. And maybe we just jump in because AliveCor, you touched on it, but I think one of the most interesting things that the company does is leveraging AI and advanced models as well as just a gigantic proprietary data set to do really interesting things, and this is one of the areas of AI I'm really excited about. I know there's discoveries that AI can fuel in medical seem really, really positive and one of the really uplifting areas of this thing. So maybe let's talk about, like, what does AliveCor do specifically around how do that proprietary data sets drive discovery and, and how does that kind of play into stronger models, and what's the balance between model and data? 100%. AliveCor creates devices such as a single lead or a six-lead ECG device that a customer or a patient can use at home, take a 30-second EKG recording- And once they take a recording, our AI kicks in, and it's able to detect heart rhythm disorders such as atrial fibrillation, such [00:03:00] as brady, such as tachy. And then the role of that AI is to be able to analyze that 30-second recording and tell the customer in that moment what's going on with their heart. Mm-hmm. The patient or a customer can take an action, such as take that recording to their own doctor, but we also have doctors that can engage with the patient. But the real genesis of that, the, the company is really just putting that technology and the capability in the hand of the customer so that they can make that determination. The business is pretty much direct to consumer, so we sell devices on Amazon directly as well as through providers. Uh, so we work with the doctors to be able to help them help their patients just understand and take control of their heart. Where AI comes in, and just to give you a sense of the scale, about 6.4 million customers globally, 400 million recordings, and as a company, we started with one or two determinations. Atrial fibrillation was the first one, and then over the period of time, as we have recorded more and more and more of these EKGs, we are able to detect more and more [00:04:00] and more of the AI determinations. We take this data, we buy data from other sources, and over a period of time, we have been able to go from a company that could just do heart rhythm disorders to structure and morphology all the way to products that can now detect heart attack. So that's the role of data coming back to us, where we can train these algorithms, create better models, deploy these models back onto our devices and technologies, and just do more and more and more in terms of determinations and detections by using AI. I think one of the really powerful things that I love about this is maybe this doesn't supplant going to the doctor or replace going to a doctor, and I know there's all sorts of claims you can and cannot make in the medical field. It's very regulated, which is probably the right thing, but being able to do some of the stuff at home and get those readings opens up the level of insight and the ability to get medical care to a much broader range of people than could ever do it before, and being fueled on the ability of it to be just more and more and more accurate. How have you [00:05:00] designed around and made decisions around the growth of model capabilities? I mean, we've, we've gone from ChatGPT-3 only a couple of years ago to these models now that break out of air-gapped clusters. But also, like you've gotten more and more data, like where do you think the value levers really come from? First thing I have to say, any model or any technology that we deploy to our customers goes through very rigorous regulatory process. Yeah. There's model, there's proprietary data, but there's also once you deploy a model, a model creates its own data paths because the customers when they record an EKG, we get longitudinal record, "Hey, I'm recording an EKG today." Typically when a patient's recording, they may take like eighteen to twenty-eight recordings in a month, depending on what product they're using, and they may keep on recording it over a period of time. So I have not just one point in time data, but data over a period of time In that, I also know what's a patient demographic, what conditions are they developing [00:06:00] over a period of time. So once you deploy an AI model, that model creates data, not just the proprietary data I have, but also the new types of data. A patient may say, "Hey, this is how I'm feeling," or, "This is what I consumed during the day. This is my activity profile for the day." So over a period of time, the models tend to get multi-modal. Like, the inputs tend to get multi-modal, not just an EKG, but a lot more that comes with it. And of course, as you're thinking through the next generation AI, the data scientists, they're always looking at, "What can I infer? What can I predict? What can I do better?" So proprietary data is one part, but I do believe that model itself, deploying that model creates a flywheel, and the generations of models get better and better and better, just because you're getting more and more and more types of signals into the model. It's really funny to me to think about how you look at this here. Thinking back to people we had on the show, we had the head of AI driver training at Waymo, and the approach [00:07:00] to training that model was really, really similar to what you said. It's get it out, get the model out, get it working, and that generates its own data, which then they can use to build inference and all those kind of things. And yes, having their own internal proprietary data sets helped a lot. But also, it was just getting these models to create their own data and kind of self-recurrently fix there. Absolutely. Even us as a SaaS company, I look at how we have evolved over the eight years I've been here, and I think the models clearly got a lot better over that time. They went from really not existing in the public domain when I joined the company in twenty eighteen to those jailbreaking air gap crazy models we hear about in the news. But at the same time, I don't think we'd be able to do the level of insight generation on your digital experiences and all of that at any kind of acceptable cost if we hadn't also built huge amounts of proprietary data around what does human frustration look like, how do we understand that, and did pre-AI data building of something ridiculous like a trillion interactive data points or something. Some [00:08:00] stupid number. Wow. Yeah, it's, it's ridiculous. But you dig into all these AI companies, and everyone has these giant numbers 'cause if you don't, you're not doing the cool things. It does kinda take both pieces. I am curious, 'cause we've kinda skirted around this issue a little bit. Clearly both data And model are vital. If we look at that kind of thought around model improvement, and every once in a while people kinda say like, "Is this it?" How do you look at that piece around, like, the AI just consistently getting better and better? Or do you kinda look more to now that you have really, really strong models, using the data it kicks off to generate further investment? Where do you put your forward bets on where you're gonna get the most productivity improvements here? That's a great question, and this is an advantage that we have as an integrated company. We do our own devices. Think of these as sensors that are creating data, and I can put more and more sensors into the products, so I can generate more and more data. And then as I start thinking about what my customers are telling me, what my doctors are telling me, what would be [00:09:00] useful, and I'll maybe just give an example. Typically, when you look at an ECG that, that a hospital is using, you may see a doctor using that ECG to detect about 100 conditions, which is great. But now there's research that says, hey, there's a doctor, maybe not a doctor, but the ECG signal in a hospital may be able to detect up to 600 different conditions. There may be conditions that required an echo testing. Mm-hmm. And then echo test in hospital will cost hundreds of dollars, sometimes thousands of dollars, takes time to schedule, and if you can do a low-cost ECG just to be able to detect that condition that you used echo for, then all of a sudden we have reduced friction. And what if I brought that back into patient's hand at home? And all of a sudden, you know, I can talk about a simple, accessible ECG detecting a very complex condition and really taking out the friction, the time, the cost from detection to treatment. That's amazing. So taking all of these inputs and then, you know, we start working with doctors on, okay, hey, what data do [00:10:00] you have in the hospital? How do we start thinking about correlating that data, patient record, and the ECG signal back into the ECG signal and all the data that we've gathered over, over years? Is there anything in that data that tells me or correlates with LVSD or LVDD, right? These are all the conditions that you detect All of a sudden, you know, we start thinking about the value. What value do we think about creating for the customer, for the doctor, for the payers? And that's how we start thinking about our roadmap. What can I do in this 30-second ECG at home that helps saves lives, reduce the time to treatment, and really with a very high quality trust signal through our DA and all the regulatory processes we can deploy to customer in a manner, in a timely manner. So everything starts with the value, the value that can provide quantifying in terms of lives saved, in terms of treatment options, in terms of financial impact, and then bringing it back into, hey, based on my data, based on all the signals we can [00:11:00] acquire, what can we do that we can take through FDA and through all the regulatory bodies that we can deploy? I wanna back up a little bit. One of the things I liked about your background and, and why I was really excited when you said you could come on, is this view of, like, the holistic business. And so it's great that you're improving the model and you're making this stuff more accessible to people, which I think generally healthcare should be, but I've had enough people from medical fields on the show to understand a lot of these concepts fail, especially medical, because the incentives across all the tiers in medical aren't aligned. The payer needs to see the benefit. The providers need to see benefit. All these kind of different levels need to see benefit, or else you could have the greatest tech in the world, but it might sadly not, not work out. How do you think about AliveCor in relation to all the kind of levels of medical delivery within the US and, and worldwide that need to work together here? That's a great question, Geoff. As AliveCor, very consciously, we started with D2C, direct to consumer. So then it [00:12:00] becomes, hey, when the patient is not in the hospital and not in front of doctor How do we serve them in a way that makes sense and give them the peace of mind? The device is something that they buy out of pocket, so the job for us is to serve that patient. While we are serving that patient, very quickly we realize we're also serving the doctor who's working with the patient, and 50 to 60% of my patient journeys start in a doctor's office. Because over a period of time, doctors realize that if the patient's more educated, and they're more informed about their health, and they're empowered, they'll respond to treatment better. So at least that level, "Hey, how you think about a doctor treating a patient? How do you think about empowering that patient to know more about their heart and take control?" That's been something that's working really well for us. And from there, the hypothesis builds up. All of a sudden, if there's value, then we can have that conversation with the payers on, "Hey, how we make the access easier?" Mm-hmm. And that's where, you know, we're working with employers, we're working with payer groups to really just articulate that value [00:13:00] and make the access easier, whether it's US, whether it's Europe, really it's the same conversation. We come in, we talk about the financial side of it, we talk about the medical side of it, we talk about the science side of it, and really just opening new doors so that today the device is not reimbursed, but maybe it could be, right? These are the conversations that we're driving with payers as well. Mm-hmm. I do wanna get into the other topic we had talked about, 'cause I just think this is so rad. I remember when the first Alexa came out, and just all the voice interfaces early on. A lot of these at the very, very beginning were garbage. You know, they were not good. We had a guy named Oji Udezue who speaks at a lot of our events, and he worked at Microsoft for many years, and talks about they spent a long, long time on natural speech and natural language processing. And before the LMs and the kind of multimodal models we have now, it was incredibly hard, and you saw this in a lot of these assistants, and they got much, much better. But I'm curious to hear about how you saw that develop, 'cause you were there for a [00:14:00] lot of the development time. What was kind of the unlocks, and then how did that ecosystem work out? Does that experience of developing those, both how they worked but also the ecosystem around them, play into how you look at driving AI agents now and, and developing the capabilities currently at LiveCore? 100%. I joined Alexa, I'm saying 2016, where it was just beginning to capture public imagination, and my job was to really just think through how are we gonna take a- Alexa into enterprise. And even at that time, there was just a lot of excitement about AI, what people thought AI could do, what people thought they could accomplish with AI. The use cases were numerous, and those use cases by and large are same today. At that time, NASA and Alexa collaborated on thinking about at some point in time, we're gonna put astronauts on a rocket to Mars. And that journey is going to be a long one, and you know, there's going to be a technical challenge, a scientific challenge, but there's also going to be just a mental [00:15:00] aspect of somebody locked in a rocket spending months, maybe years in space. And how do you keep them sane? How do you keep them engaged? How do you keep them productive? And maybe Alexa can help. So even like thinking through that use case and bringing back into, hey, what does an astronaut do? How do they live in that confined space? What are the mental challenges? What are some of the things that they would wanna do? And really thinking about short of piloting a rocket, none, none of us wanted the hull Uh, open the pod door scenarios. Like, but really thinking about, hey, what are the mental aspects? What are the challenges an astronaut is going to face? So really thinking about those use cases, making it into a product that is millions of miles away in a very constrained scenario. These were some of the use cases that were coming up. We did some experiments with NASA. Alexa was part of the Artemis I mission that went around the moon, and we did, like, a lot of very interesting technology demos. Similarly, there were explorations about Disney [00:16:00] thinking about their guest experiences in the Disney park. If somebody's coming in and, you know, they're gonna spend time in the park, how do we make that experience more enjoyable? How do we make sure the time that they're spending in park is to their liking? So thinking about the use cases, what could voice agents, what could Alexa, what could a Disney agent do? And then solve a lot of really hard problems that are still relevant today. And I think over a period of time we'll see, hey, I have agent one and I have agent two, and these are not isolated worlds. They're gonna have to work with each other. How do you think about agent to agent? How do you think about handoffs? How do you think about different capabilities? How do you think about just that seamless customer experience that you can deliver through agentic AI to an end user, maybe enterprise, it may be consumer. But really just stitching together that customer experience that's magical. Yeah. So I guess as you look at that, like, where did those go forth? If you look at Disney, and maybe this, you know, I know Disney has more NDAs than, than I would probably like to think about, [00:17:00] but I've gone to Disney with my kids in the past year, and it was a magical time. How did that play out where you're trying to take an already magical thing and, like, what can you do or how can you improve that in this already ecosystem of cool stuff? Working with Disney, there were, like, two different pieces of experiences that we deployed. First one, and I don't know if that still exists in Disney rooms in Orlando, but first one was really just a Disney branded guest experience in a-- is in a Disney room. So if you go in a Disney room, right, you would have, like, a Alexa device that was built as a Disney device, and you could do, "Hey, Disney," or, "Hey, Alexa." That was meant to enhance the guest experience. Mm. So you could think, ask about your reservations, show timings, take certain amount of actions using that Disney device in your room. Mm-hmm. And second was more of a take home experience. And this is, again, I don't know if that's still available now in production. That experience was, hey, I went to Disney. I had a great time. Can I take a subset of that experience with me to my home? Mm-hmm. And [00:18:00] that was a Disney skill that Alexa launched on Echo devices which could mimic Disney characters, which could bring, like, you know, Disney News and Disney entertainment to Alexa devices at home in a very certain manner. And again, this is all built on Alexa. This is a previous generation technology. A lot has changed since, and I haven't kept up on what happened to those experiences. Thinking about the learnings here, right? Like, so obviously start from what people want and how you can deliver and just look at the tech limitations. I guess, what lessons from this era carried through to now, but also maybe more interestingly, like what are the things that were learnt then that people really need to unlearn now that might hold you back from success in the more advanced capabilities of these current models? Let's maybe just start with what are some of the durable things that were learned and we learned, and we should all keep in mind. Every technology, every business, every real piece of software or AI that we deploy Has to have a purpose in mind. Who's going [00:19:00] to use it? How do you think about that user experience? How do you think about the friction, the frustration, and how do you go very systemically about eliminating that? And while you do that, how do you make sure that it's a sustainable business over a period of time? And maybe there's some lessons from Alexa there that we'll talk about. So that's number one, just keeping that customer at front, right, and center, even though when you are at the leading edge. When ChatGPT came out, it took us a while to understand what can it do. And once I got comfortable, we did more and more and more. So really just deploying that technology in a very human way so, you know, you're not just leading with the technology, but really leading with, "Hey, let's get you comfortable using what it is. Let's get you comfortable understanding what you can do, and then do more and more and more." Second thing that I cannot emphasize enough, and these are some of the hard lessons learned over the last 15 years, is technology's important, customer experiences are important, but the business model and the financials are equally important. Really just, you know, what you were talking about. [00:20:00] We deploy technology in Disney with the goal of creating an experience and a business model. So you can start thinking about, "Hey, if I'm creating a piece of technology, how do I think about ROI on that? Is that going to be sustainable? Can I keep on funding this business out of the ROI and the economics I'm creating?" Because at some point in time, music does stop and you run out of runway. So really thinking about at least a hypothesis at the beginning of what is it that customers are willing to pay for, and how do I think about the value? And I think potentially even more so now than ever, because everyone's kind of talking about with progresses in AI and, and the models, and the ability to build faster and faster, coding is no longer the limitation. I would argue almost at no point was the time to code something really, really the limitation. Maybe a little bit, but there was always a function. You could go hire more engineers probably. But it was... there's only so fast your customers can ingest- Yes ... new things. And I think we had almost a rate limiter [00:21:00] at one point because you did have that kind of engineering rate limit. But now if you think about the core thesis of what drives that, it's still the same thing. You need to pay for the stuff. You need to figure out how the economics work, as well as what you're producing to create a great experience. You can create 10, 20 things that might all be wonderful and magical if the users use them, but they can only learn so much. No one cares about your product as much as you do. So 20 great things might all sit there flapping in the wind if you don't have the other pieces of your business, distribution, uh, how do you drive adoption, all these things. The product has to be marketed. That's right. Yeah. So what I'm hearing is a lot of this stuff, the lessons go round and round and round, and the tech changes, but there's the core lessons to be held onto. I guess the negative lessons I figure are probably more concise than global and forever. Are there little things that we need to forget from the days of building the early Alexas and things like that, that might impinge on building great tech and, and use cases around the modern LLMs now? I would say, hey, one of the limitations on Alexa was the business model. So we talked about that. Yes, [00:22:00] we could sell the devices, but how does the usage make money? Especially at that time we were dealing with NLU, the technology, the inference was not that expensive, and all of a sudden you are dealing with very real cost of inference. So how do you think about making sure you and your customers are aligned on the business model and really thinking about what that looks like over a long time rise. Other than that, really just sometimes, you know, we can get ahead on our skis on the technology capabilities, get super excited and deploy things that may not deliver value to customer either in that moment or much later in the lifecycle. So really just keeping that customer in mind and where they are and the pathways to market. On the positive side though, I do think that the bigger lesson, and that's been the theme, is the flywheels. As we think about investing, right, you can invest in a product, you can invest in a feature, but the best investments that I've seen are in a flywheel where you say, "Hey, I can build a better model." That better model gets me more data, that helps me get out [00:23:00] to more customers, that help me deploy my technology, make the technology better incrementally over a period of time. So I think that's been more fundamental lesson. Like don't look at a product, don't look at a feature. Really think about if there's a flywheel behind it. Mm-hmm. What does that flywheel look like and how do you make sure you're making the right investments, not near term, not even midterm, really thinking about a flywheel that may play out over five, six years. It seems like so much of what we talked about were acute questions that had big, longer, more durable answers, which I always find really cool to see kind of over time. You know, I always kinda joke, marketing and product and all these things haven't really changed in a thousand years, right? It's make something that people actually need. Find out how they can get it. Where are they in a mental place to think about the problem and, and the solution. Go there. Tell 'em about it until you want to puke. Make sure that they can have a low friction way to acquire the thing that's valuable to them, that creates more value for them than the cost of you making it. We got a business plan. It hasn't changed since the days [00:24:00] of finding a sharp rock that you can trade for something else. It just, the underpinning machinations maybe got more complex, but like a lot of the best tips I've found are the ones where we just apply those in smarter and smarter, more finessed ways. Rajiv, I gotta be honest, it was awesome having you on. I could go on talking to you for probably another couple hours, but I think it's time I gotta let you go. But if people are, you know, wanna reach out, ask questions, just even say, "Hey man, that was spot on. Really cool stuff, Rajiv," is LinkedIn the best place to reach you or is there a better spot people can reach out? LinkedIn is the best. I'm also on X. Awesome. Thank you for coming on. It was a real pleasure. I hope we can stay in touch and look forward to talking again. It's been my pleasure. Thank you so much for having me. Thank you. Have a good rest of your day. Thank you, Jeff.