Tanay Tandon on The Robot Brains Episode 5 Season 2

Edited for clarity

Pieter Abbeel: Today, here with me is Tanay Tandon. Tanay is the founder of Athelas, a startup for analyzing a person's blood. The cells within are often used to diagnose many health conditions and illnesses, from infections to leukemia and bone marrow disorders. Generally, it's a long and expensive process. You have to go to the doctor, have a sample taken, and wait for a couple of days for a trained professional to then analyze the blood. And finally get your diagnosis. Athelas is using machine learning to dramatically improve the speed and efficiency of testing blood cells. From a simple finger prick worth of blood, Athelas devices can monitor and help healthcare professionals to remotely care for patients. Today, over 30,000 people have used the health devices to monitor a range of conditions, including hypertension, and diabetes. I'm really excited to have you here with us today. Thank you for joining us.

 

Tanay Tandon: Thanks for having me. We're super excited to chat through just some of the work that we've been doing in the area.

 

Pieter Abbeel: Well, it's really exciting to have you on. 

 

Tanay Tandon: Yeah, definitely, I think one of the really exciting things about working at a healthcare business and, you know, working on healthcare problems as engineers, as salespeople as operations people, as if there's this broader mission. That is very tangible. I mean, you look at healthcare as an industry. It's the number one reason for personal bankruptcy in America. We spent $4 trillion a year on it. It's one of the largest drains on the economy. When we're thinking about the end state and really the mission for dollars, it's it's that, in my opinion, and in the opinions of a lot of people. There’s about three things in the American economy that are fixed. They would completely change the way that the average middle class American would start living like royalty. And I mean, there’s healthcare, which is a black hole of wealth often. You know, if you go to a hospital, one, suddenly there's this $40,000 bill. Also at education, you know, you look at the student loan crisis, you look at the cost it takes to get, you know, standard education in America. And then real estate, I think that the cost of rent and the cost of actually living in a place just, you know, the basic action of existing is so expensive in America. And I think that, you know, we're hopefully taking on that first problem of healthcare. Which is how can we make healthcare lower cost? How can we use technology to make that happen? And so, in our welcome packet, there's just a really short note that the leadership team and I wrote that we put in a box along with the sweatshirt and a jacket that we give people. The note just talks about, hey, like you're about to start working on this problem with making healthcare more preventative and solving healthcare in America. And I think it does get people excited. It gets me excited to think about that as a mission that we're growing a team around every day.

 

Pieter Abbeel: Can you say more about it specifically? Why did you decide to look at blood? And specifically, the things you’re doing at Athelas. I mean, you started this many years ago. What inspired you to go for specifically what you're going after right now?

 

Tanay Tandon: So it's an interesting story. My co-founder, Deepika Bodapati, and I would compete in high school science fairs against each other. And in the Bay Area, science fairs are very, very competitive. People are working on projects that are, you know, well beyond, you know, my comprehension. Her work was in imaging and, and molecular imaging in particular, and trying to look at the early signals of cancer. My work was building very simple natural language processing models, or in some cases, computer vision models, to do basic tasks. Like one year, I built a summarization engine. Another year, I worked on a simple tool that could classify Malaria cells that actually ended up being the precursor for Athelas. 

 

But the reason that, you know, we started this together, and the reason that we started the company and inspired to keep going, is, you know, when you look at healthcare as an industry. In particular, something like 60% to 70% of the costs are in labor, and are actually in the services component of health care. So when you get a blood test, you know, I was shocked to find out that, like, literally all the all the costs of that, like the technology to run the test cost a couple pennies, the time on their system cost a couple pennies, and then the time of the operator that’s pre-processing the blood and then in some cases, analyzing it under a microscope. And, you know, manually labeling cells or, you know, writing things down. That's like the $80 that you're paying for that blood test. 

 

And so you see this again and again. Pharmaceuticals, you know, I think politicians…not to get too political…but politicians will talk about the pharma industry all day night, but pharma drugs spend is 8% of healthcare spend, whereas physician time and nurse time is something like 80%. And so, you know, when you think about where the dollars are going, it becomes very clear that, okay, well, we want to tackle problems where we can augment a single physician or a single nurse, and make him or her more powerful in order to scale to more patients at the same time. And the way to do that is technology. And I think that when we started, the goal was always to build a lot more types of tests and a lot more types of sensors and take on more and more of the market. But we wanted to start with something that we knew we could build and was technologically possible, and also was a large enough chunk of the market. So the CBC (complete blood count) is where we started. It's the most commonly ordered blood test in America. Like you mentioned, it can help detect anything from an infection all the way to leukemia. It really is considered that first wall of a diagnostic. And you know Deepika, my co-founder, who is way more familiar with healthcare really identified this as the test to go after, because of the breadth of conditions that could help solve. And also the fact that you'd seen how it was running a lab where it was an automated cell count, and using impedance to track each cell in a large $100,000 system, combined with a pathologist sitting on a microscope. And both of those things seemed like things we could automate with computer vision, and some basic microfluidic. So we just started running at that problem more out of curiosity to see if we could build it, rather than actually having identified a good market and a good business model. It is probably the wrong way to start a company. But we eventually iterated our way into something that now serves a lot of patients and is quite useful.

 

Pieter Abbeel: So CBC, complete blood count, what exactly is the result of such a test? What does it have to do? And how are you able to do it so much cheaper than the previous machinery?

 

Tanay Tandon: Yeah, definitely. So the test that we offer today, white blood zone neutrophil counts along with blast cell detection, and we also detect a ton of other types of white blood cells, platelet clumps, etc. So a good chunk of the CBC and this quarter, we’re launching platelets and hemoglobin, which round out the rest of the CBC. What you can do with this test essentially is it tells you the number of white blood cells in your bloodstream and the types of white blood cells in your bloodstream. And why this is really useful is let's take an example of a patient that's on chemotherapy during cancer. When they are receiving chemo, they are receiving a cytotoxic. A chemical that's essentially a drug agent that is killing a lot of the fast growing cells in your body, which include tumor cells, but also include some really important cells, like neutrophils, which are a type of white blood cell that are an infection fighter. And as a result, these patients become immunocompromised. They can get an infection that might kill them. They can get an infection and put them in the hospital. And so measuring and tracking that neutrophil count is really important in those populations. Similarly, you know, if there's a patient that has a high propensity for internal bleeding, or, you know, post operative after a surgery, You want to make sure that there's no massive inflammatory process that's going on. You want to see a good recovery so then you'd want to be measuring platelets and hemoglobin and also white blood cells on an ongoing basis. And so what we've done is take a microfluidic tester of cartridge. So with a small volume of blood drawn, we create a monolayer of cells. So that's a layered single cell thick. And then that test strip, it's consumable, is inserted into our device, which is about the size of an Amazon Alexa. And then it takes hundreds of images and classifies them using convolutional neural nets. That first segment where all the cells are, and then classify the types of cells that are present. So like, you know, pretty standard. It took a lot of time to train and figure out, you know, how to optimize but at its core, the model itself is not something fairly complex. And then after that, we have a count. So hey, these mini cells per microliter, microliters, this type, and then percentages of the different types of white blood cells and platelets and hemoglobin. And from that, a physician or nurse, or any clinician really can can take that number, and make a determination as to, ‘hey, this patient’s counts are really low, let's give him or her a prophylactic antibiotic, which is, you know, precautionary antibiotic, because there's a good chance they might catch an infection.’ And previously, this process would be like a three day turnaround. You would go into the lab, you would get your test result, and you'd wait, the results would come back. Your doctor would call you, then you'd have to call back in or something. You know, in our case, it's two minutes, it's right there, if you take the fingerprint, the result happens in front of you. And I think number one, why it's cheaper is that the intelligence aspect of it has been completely automated with computer vision. And number two, we've used off the shelf hardware and really simple hardware tools, to essentially take what used to be this big $100,000 piece of equipment in a flow cytometer and then use software to instead supercharge it, essentially, with a simple optical microscope to classify themselves. And so I'd say what we've done is drastically simplified the physical components. An autometer might have something like 400-500 parts in it. Our system has like four and, and instead, all the other components like the light based impedance and the tubing and all those things that go in the postometer. We just trained that intelligence into a computer vision model. And so that's our core device, Athelas One. And then along with that, we've since expanded and we have a suite of other tools for blood pressure monitoring, for detecting hypertension, and weighing scales and glucometers. And we built simple sensors for hardware for pill tracking for medication adherence. Our belief is that with these suites of sensors, we can start getting a better picture of a patient's health and preventatively monitor them. And then use simple machine learning models to spot trends or you know, detect blood cells or whatever it may be to, again, augment the intelligence of a physician or nurse.

 

Pieter Abbeel: I love how you're describing this because if I heard it correctly, it means instead of having blood drawn, I just prick. I just do a little pricking my finger and then I swipe the blood into your device and it takes care of everything else. And I can do it as many times as I want. Is it costly to run the device?

 

Tanay Tandon: Yeah, exactly. I mean each test trip costs us about $1 to make and manufacture. But other than that, it's all software and so we've drastically increased the frequency at which you can run this test. And so it's a pretty big deal for the for immunocompromised population, because in that sense,

 

Pieter Abbeel: It not just changes that you can do things more easily, it also changes how often you're willing to do tests, I mean, getting things scheduled is always a real pain. And now you just do it on your own time. That's amazing. Now, one thing I'm curious about on the AI side of things. You're saying computer vision is used to look effectively at a drop of blood. Somehow images are taken of that blood drip, and computer vision makes sense of it. Now most machine learning these days, maybe yours, too, will require labeled data to do the thing on its own. So where do you get your labeled data from? What's the ground truth for the systems? You train them?

 

Tanay Tandon: Yeah, definitely. I'm a big believer in ‘first try to do the thing without machine learning.’ And when you actually need to use it, you figure out what the hard parts of the problem are, that need a ton of data then hone in on those. And so when we first started, I mean, we were able to, you know, use pretty standard image processing. You know, techniques like transformers, and edge detection to classify cells and segment them out. And then like, to count and sort each cell and its nuclei in it. And based on the shape and sizes of the nuclei, you can classify the exact type of cell, whether it's a neutrophil, which is an infection fighter, or a lymphocyte, etc. And even for that, we were doing basic, you know, edge detection, and, you know, simple transformations, and then counting the number of nuclei. And that worked really well in a simple controlled environment. Now, the second you start having these edge cases like the cells now might be of a slightly different shape, or slightly different size, or the ambient lighting is a little different. Suddenly, that starts breaking down. And that's where the neural nets that we need to train have to be so robust to all of these. These perturbations and edge cases in the different cell types. And so what we did is, is that we worked with pathologists. We work with pathologists on contract from a bunch of different institutions, and we created our own label dataset of hundreds of 1000s of these cells, and then obviously we use the usual tricks of image augmentation and, whatnot, to really multiply that data set. It would work particularly well for cells, because often the transformation you're looking for is a simple rotation or the cell is slightly more often in one direction, because of how it appeared in the image. And so, you know, with with a couple, I'd say, I'd say like, 20, to 30, pathologists, and ourselves also working on the data cleaning and data and process, we sort of created this massive dataset. And then we're able to train the neural nets on it, I'd say one really important thing is, is actually making sure that the, for example, if we just use a training set from someone, you know, looking at these cells under a microscope. There would be some inherent differences then the images collected are our device, which has resolution differences and lighting differences. And so the training set had to be built on the devices that are going to be actually used. And so the exact lighting, the exact imaging, and the exact resolution. That was a challenge. But you know, it took us over the course of a year. We sort of built this production grade computer vision model, and also this production grade data pipeline to go with the blood labeled data with pathologists. Early on in a computer vision or machine learning startup, having production grade data pipelines is equally if not more important than the model itself. That's something we learnt the hard way, but we got through after a bunch of iterations.

 

Pieter Abbeel: Now, this is one of the things that's often in the news is how some machine learning systems might work well on some subset of a population, but not on other. For example, face recognition systems were notorious to be less accurate for black people than white people. And so I'm curious…in blood, are there any such concerns? Like is it possible that some populations have different types of blood cells and at least in the way they visually appear? Would it escape the accuracy that you generally achieve?

 

Tanay Tandon: It’s a great question and a really important question. For example, African American populations will have a higher propensity for sickle cell anemia. Genetically, it is a condition that's far more prevalent in African American populations. And, hats off to the FDA, because in the review process with the FDA, these are the things they really hone in on. They want to see samples from a wide range of demographics. They want to make sure that you're both training your sets, as well, as you know, more importantly, when you're, when you're actually testing the system, and, you know, validating it. You're running it across patients with a wide range of diseases, with a wide range of, you know, ethnic backgrounds and age and, you know, all those things. And the other pieces that they really look for are certain rare diseases that might impact the way blood looks and the way blood cells look. And so they, I mean, when you're in the process, it's honestly quite annoying, because they'll mention some rare disease with which there are 20,000 people in the world that have it. And now you have to go find 100 samples of that, and, you know, make sure your model is able to perform on it. And so whenever you get that response, and you know, the FDA will say, hey, we need a patient with, you know, this genetic disease. And like this blood disorder, it's a pain because you have to go hunt that down and, you know, work with labs and work with clinical facilities to find samples, matching that background. But it's really good from a public health standpoint, because it actually ensures that your system is robust to these edge cases. And I think, for us, and I think in general, for you know, as technology companies enter healthcare, that's one of the most important things is really ensuring that, that the models that we build, and the models that we validate are robust to the full gamut of the population that they're going to be serving. So yeah, that's definitely something we had to work on in a big way.

 

Pieter Abbeel: Well, that's really impressive what you're doing there. And I'm also really impressed with the FDA, a government organization that actually gets it right and understands how the machine learning system, or other systems also, of course, need to be tested across a wide range of populations to be able to achieve the desired outcome. Now, one of the other things that, of course, comes up a lot these days, especially with the thoroughness trial happening is well, wondering, what's the difference, right? I mean, Theranos, was a company that promised from a low prick, you can get everything tested. And here you are actually delivering something similar. But what they were claiming wasn't right, and your thing works. What's the difference? Can you say a little bit about what's the difference between Athelas and Theranos?

 

Tanay Tandon: Yeah, definitely. So I think number one is, are the tests that we have, you know, we've worked on whereby the CBC is a very specific test, it's not 30 blood tests, it's not 101 tests from a drop of blood. It's one or two core blood tests that are very important and highly utilized. And so step one, I think the fact that we're really focused on this on this one test is distinguishing. And then number two is, is really the focus on a strong clinical review, FDA trials, working with peer reviewers, with institutions that publish. And really just honing in on the data, you know, all of our data on the devices performance is available online. The way the device works, it's public. There's a lot of people that when they're building technology, there's an attempt to obfuscate or an attempt to make it seem more complex than it really is. But I think the beauty of what we're working on is that it is simple, and it is actually fairly easy to understand, which is, you know, you take blood, you create a monolayer of it, and then you count cells. And I'd say one of the more important parts is that this is the way that blood has been, you know, analyzed for decades, if not centuries. You look at it under a microscope, and then you count the different types of cells. All we're doing is automating the intelligence portion of that. And so when you look at the risk factors associated with this specific type of test, it all comes down - can you accurately detect the various cell types that are present? Can you automate something that humans are already very good at? Which I actually think is a great, great example of a problem that machine learning can solve. We work with the FDA from the earliest stages, you know, we meet with them a couple times a year or, you know, have conversations with them at least a couple times a year. So it's super important to us that our interactions with the broader clinical community are one of, you know, we want to work with them. And we want to make sure that when the FDA asked for a trial, I don't think we pushed back once. It was about if we would run the trial and the work. We collect the data, we would show that it works, if the worst performance limitations, you know, in the labeling in the conversations with the FDA, those are made transparent and very clear. And so I think just focusing on the science and focusing on the data, was one of the ways that we've tried to differentiate ourselves, and also scoping the problem to something that scientifically can actually be done because humans do it every single day today.

 

Pieter Abbeel: I mean, it's fascinating. It works. Now, if I'm somebody who wants to start using your system, your devices, where do I start?

 

Tanay Tandon: The system itself is indicated for prescription use only at the guidance of a clinician. So we work with specific chronic disease populations that really need this type of monitoring at a high ROI on a high frequency basis. One example is there's a refractory schizophrenia drug called Clozapine. It's used to treat schizophrenia, and it's very effective, but it has a certain side effect, by which in a certain percentage of patients, you do actually see a drop in neutrophil count which makes them more at risk for developing an infection. And so there's a population that our device has been incredible for because until today, in order to receive the medication, each patient needed to get a venous blood draw once a week. And our system has transformed that into a finger prick. And doing a finger prick once a week is way different than, you know, doing a venous blood draw once a week, just from a logistics and operational standpoint. So it truly has been life changing for a lot of those patients. 

 

We talked a little bit about oncology, you know, with chemotherapy in cancer care. So, you know, today we're really honed in on delivering great care to those populations. I think the end goal in maybe a year or two years is for us to have enough tests and enough ready to go analytics in the device that it's useful for the general consumer. And it's something that, you know, you can buy online and have at home. You know, it's a prescriber or physician connected device, where after a result, a quick console to parse the results and ensure that you're, you know, understanding them correctly, is in place and that's the goal we're working towards. How much of healthcare can we transition into the home? And how much of healthcare can we make preventative? If we can, we'll start bending the cost curve and really drastically reducing costs in the US healthcare system. So that's what you know, today as a patient you would have to get prescribed by a physician, have a specific condition then that physician prescribes it for and then you get up and running on the system. Hopefully in the near future it's something that anyone can just order online

 

Pieter Abbeel: So I actually read a story on your blog, which was really a great story. I recommend people check it out. It is about a patient whose life was transformed by this, right? I'm curious how there are other diseases aside from schizophrenia that it's used for?

 

Tanay Tandon: Yeah, definitely, um, we have hundreds of cancer patients today who use the system in their home while treatment in order to measure for immunosuppression and dose adjust. In autoimmune populations, if you know it's a patient that has rheumatoid arthritis or another condition. That's a population that we serve in a big way.

 

So in autoimmune populations, you have inflammation which is the underlying condition and you're receiving drugs that are immunosuppressive. So it's similar, you know, to the Clozapine use case, in some ways that you, as you're getting these infusions, you want to be tracking your white blood cell counts. Number one, you want to make sure they never go too low because of a risk or infection. Number two, you also want to use it to track the progression of inflammatory response, which is like your white when when there's an inflammatory response in your body, a couple of metrics go up, there's a ratio called V and LR, which is neutrophil to lymphocyte ratio to the strong marker for inflammation and something we can measure or the you know, just they've been total number of white blood cells. That's something that's really useful in autoimmune populations. 

 

There’s one story that I don't think is listed on our website, but probably the most important story for the inception of the company is when we detected leukemia in a patient in our very first clinical trial. And you know, leukaemia is a liquid tumor. It essentially manifests as a massive proliferation of white blood cells and, you know, immature blood cells in the bloodstream. And so it often stems from an underlying bone marrow disorder. And the, you know, I remember, this was the first time we'd ever use a device in a clinical setting, we were just trying to collect data, right towards the end Y Combinator, our goal for the summer at YC was to was to have a system that was functional and you know, passing a basic validation trial. And during the trial, Deepika, my co-founder, was one that was picking every patient and I was sitting in the back room, running the tests and running the samples. And one patient came by and she got the sample. I ran it and immediately it was something like 40,000 white blood cells per microliter. The system classified that and on top of that, the cells were also very large, which was detecting blast cells, which are immature cell types. I knew this was leukemia. Obviously our system wasn't FDA cleared at the time, so I couldn't, you know, go tell a patient. I don't know, I'm not a clinician.  So we let the nurses know. And they followed up and three weeks later, the patient was confirmed to have leukemia and I think for us, this was one of those mind blowing kind of moments where we were like, Okay, we built this thing. Its a Raspberry Pi and wires, everything's dangling out of it, and barely barely works, you have to like hold it together, but it detected leukemia and that to have detected leukemia after a patient had visited her doctor and had that, you know, consultation, but the diagnosis had gotten missed, because you know, today's healthcare system is not particularly quantitative. I remember on the flight back home, Deepika and I were talking and we decided to take a leave of absence from school and go full time on this because it became very, very clear that there was a massive market here. We may not have figured out exactly how to, you know, nail the business side yet, but healthcare wasn't preventative and the fact that it took us to detect something as severe as leukemia I think was telling in the fact that the system needed more preventative technology and more ongoing monitoring technology. Leukemia is another example of a condition that can be detected and monitored using our device and is used for that today.

 

Pieter Abbeel: And I'm curious, do you ever see the full story so somebody gets diagnosed, but your device can be used to also monitor them? And, you know, you could hopefully see that your device also is helping them recover, and so forth? Do you ever see that? Or are you far removed from the recovery stories?

 

Tanay Tandon: We definitely hear stories of you know, patients but usually we’re not involved in the detection pathway today. So those are more one offs because our system is deployed more for ongoing chronic care where the patient has already been diagnosed. But we do hear incredible recovery stories where a patient will, you know, use the device over the course of several months during their chemo during their, you know, during their treatment, or on Clozapine. They use it every single week, and then the patient gets better and better and better. And that's, I mean, honestly, the most exciting thing where, you know, you see patients with for example, schizophrenia patients who are you know, a few months ago, just I'd like on the streets in some cases, and then over the course of their treatment with his medication can now even hold basic jobs. Their basic needs are met and they have a clinician and caregivers taking care of them. And they're receiving their weekly blood test through Athelas for $1. And one of the things that we've seen across the 1000s of patients that we serve is, when you introduce, you know, Athelas for monitoring, you have a 60% boost in the utilization and adherence to medicine. It's way easier to take a fingerprint than a venous blood drawn. And so it's nice to see those cases where we actively help the patient stay on the medication, and continue to use the medication and benefit from it, and then see that recovery curve for them.

 

Pieter Abbeel: So that’s where you're at today. You started this several years ago, in fact, as I understand and you well, you started a company at age 15 called Clipped. And then at age 17, I think you started Athelas. Most people are still in high school trying to get ready for college, things like that. What happened there? Huh? How did you start a company at age 15? And then the next one at age 17. Can you say a little more about your background there?

 

Tanay Tandon: Yeah, definitely. I would say that number once, I am super lucky to live in a time and place where you can just work on these types of things. The fact that you can go on GitHub and start installing NLTK, the natural language processing toolkit. It was able to do some pretty interesting things in a couple hours. And so Clipped was one of the first apps that I built. The whole idea was very basic involving techniques around named entity recognition. Like sentiment analysis, off the shelf libraries, I was not training things of my own at the time. And so I built this extractive summarization model, where it would pull out little infographics or it would pull out sentences that were core, and you could get a gist ideally, of the article and in a couple in a couple quick seconds, as opposed to reading the whole thing. 

 

And I remember I would use it when preparing for debate cases. And so it was a quick way to like, pull out the core numbers. And there was like simple pattern matching, you could do like, you know, number and then like the prepositional phrase that comes after it, and then like pull that phrase, and then you can get like, you know, 50% of patients have this issue and like and you know, it was it was a fast way to consume information. And personally useful. And at the time, actually, I was really, really fortunate because I was looking for a summer internship and I reached out to do a bunch of folks at Stanford and the AI Lab about this early thing that I was trying to work on. And I got one response, or I think it was one or two responses. And one response was from Richard Fikes who said, okay, cool. You can come intern here for the summer. And he gave me a task to work on and I bungled it. I did not do a good job on it at all, but he still gave me a shot. And so I was really fortunate. He took an early bet on me when I was still in high school. And I learned so much that summer, just absorbing just the work of very intelligent people in in that lab. And, and then I think while I was there, or you know, a couple months after my internship ended, decided that there was this this really interesting problem in in, you know, applying these computer vision models, to things beyond self driving cars like facial recognition in healthcare. That's when my co-founder and I would talk about it and we sort of kicked things off. I'd say we left school right after freshman year, right at the beginning of sophomore year, but I've been working on some of the research ideas in high school as well. They only really came together as a company in the last few years.

 

Pieter Abbeel: Where did you meet your co-founder?

 

Tanay Tandon: We grew up in the same area around here. And while I was interning at the AI Lab, she was working at the Multimodality Imaging Lab which is just across the street at Stanford. Her work was in molecular imaging, and I knew her from science fair projects. I was super familiar with her work and we became good friends. And then when it came time to start something, honestly, the ideas around Athelas were like the perfect confluence of research in imaging and machine learning, and so it seemed like a very natural progression to try to build this thing together.

 

Pieter Abbeel: Yeah, join forces with your biggest competitor from the science fair. Now, it still must have been a bit of a conversation, though, because, I mean, normally people go to college for four years. And sounds like you decided after a year to start a company. You must have had conversations with your co-founder, with your parents, etc. I mean, what was the dynamic around that?

 

Tanay Tandon: Yeah, I mean, it was interesting. Well, first, Stanford makes it very easy to come back. And so that was the core of my pitch was, like, ‘Hey, there's this opportunity here, like we're doing Y Combinator this summer, Sequoia just offered us a seed round. And I just have to go for the shot and really spend a couple months and dig into this and see what it can become. And I would say that the conversation came down to that this is an important problem to work on. And number two, I think we have a shot at actually building something here. And number three, if everything doesn't work out, then I can always go back to Stanford, they're very accepting of returning, you know, years down the road. That was helpful. It was still a difficult conversation, because the path that we took was just very different. And you know, all of our friends were in school. Honestly, there were days when I was kind of like, ‘what are we doing?’ Especially when you get stuck working on it during a clinical trial, or, you know, during a bench trial, but I think it got easier when the initial couple hires that we brought onto the team were just incredible people that were smarter than us and much more mature and just did an incredible job of taking this from a hacked together science project or idea into into a real company. And so we're super fortunate that we got some of those incredible engineers around the table, and many of them are still working with us today. So yeah, I'm glad that we took the time off from school to go do it.

 

Pieter Abbeel: It's also my understanding at that time, you were the youngest one going through Y Combinator. How was that experience?

 

Tanay Tandon: Yeah, at my batch at Y Combinator, there were a couple folks that were around freshmen and sophomore in college. And so it was cool, because I think YC is such a great environment for that kind of energy where you're figuring it out. A lot of the friends that we made during that batch, we're still super close with and so really lucky to be working with such smart people. I'd also say that going through YC, at that point of time, it was so important for the business, because it paced us. When you're working on your own, when you're outside of school, I think one of the biggest challenges is, ‘well, how do you know what goal to set?’ And how do you know what metric am I going to measure myself on week over week in terms of performance? Especially when you're not the consumer app that's launched to the public. We really benefited from that structure. We benefited from someone being there to tell us, you know, you need to accomplish these things in a week, not a month, and it just really pushed ourselves to go make those things happen.

 

Pieter Abbeel: I'm curious, from all the experiences you've had. For other aspiring founders out there, is there some advice you might have for them to pay attention to? Especially in the early stages of trying to build a new business.

 

Tanay Tandon: Yeah, definitely. I think the most important thing is giving yourself enough shots on goal. What that means is that oftentimes folks will raise a couple million dollars or a couple $100,000 and that money will burn quick. There's a number one reason for, you know, companies dying, and either the co-founders fighting, or the company's just running out of money. And what we tried to do is give ourselves enough iterations and enough shots on goal to really figure this thing out. And so iterate really, really fast like time is the second most valuable resource where we tried to push ourselves to experiment and learn from it. So that might take a month but we tried to do it within a couple of days, and so that we could quickly learn and then pivot and like, you know, change direction as needed from there. And like, when we started, we really didn't know how to build a business. We had some sense of how to build a technology, but I think we just gave ourselves enough time where we learnt the other pieces. And so that was super important. 

 

The other advice that I'd give is focus on problems that you have a sense for like the end solution. And what that end workflow looks like. For us, it was so clear, because we knew, ‘Okay, if we make this into a fingerprint, it's going to improve the lives of these patients, and these patients and these patients, we now just need to go build it and get it cleared by the FDA.’ And obviously, the commercialization was a lot more complex than we were originally thinking. Everything is more complex than you originally estimate. But really, we were able to sprint at one core goal for those initial years. I highly recommend building a simple goal to really just go after, and then go sprint at it. Focus is one of the most valuable resources a startup has. You only really have to do one thing well, and then everything else will fall in place. And then you grow from there. So those are some of the things I think we learned along the road.

 

Pieter Abbeel: I like these lessons. And another thing I'm curious about is it seems like one thing you've alluded to is that you build up a great network of other founders, and you can exchange ideas and lessons as you go along. What do you think today would be the best way to build out such a network if you're just getting started?

 

Tanay Tandon: I mean, one of the best ways to do that is just build stuff, ship some code, and then see what open source projects are going on related to it. I think the other way is, like, create a quick MVP in some project that you're working on, try to get friends to use it, and eventually, somehow it finds its way to interesting people. 

 

I'd also say that Y Combinator truly is so helpful. The batch sizes are like 200 companies so it's obviously really hard to get into but we applied to Y Combinator like three times and it's one of those things that really changed the trajectory of the company and introduced us to a broader network. I would highly recommend that. Its a great place for first time founders to quickly supercharge your network. So yeah, I think those are some of the easiest, easiest ways to build a network is to contribute to some open source community, build  early versions of a product and get it out to people to get them to use it, or at least criticize it. And then find communities like Y Combinator that are like startup schools. And I know they host a lot of events and conferences, both virtual and in person, but they're really high leverage.

 

Pieter Abbeel: Well, thanks for sharing that today. And now, I'd like to go back a little bit to your bigger vision. Which I mean, ultimately, I imagine it's part of why you're able to execute what you're doing. You have to pitch a big vision to your potential funders. Where can all this go? How impactful can this be? And I'm curious, when you look, let's say 50 years into the future. What do you think Athelas could be?

 

Tanay Tandon: I think there's going to be a care delivery platform in the home for all of us, healthy, chronically ill etc. There's going to be devices and sensors and software that we need to transition healthcare into the home. And there is a massive business to be built doing that, and in parallel, really reducing the costs of healthcare in the country. And so I think like in five years, if things go well, we will have a suite of blood tests, we have a suite of sensors, you know, some of the things that are already mentioned around, you know, other types of physiological monitoring. And we've successfully paired that with physicians and successfully paired that with, you know, medication delivery, and a fleet of nurses. We already have a team of 30 nurses who are responsible for the results of patients and coordinating between their doctors, and working with caregivers. We’re able to bill for some of those services, insurance or work with a, you know, physician to bill for those services under insurance. And so the, you know, I really think that what the end state looks like for us is, we have this incredible package of technology products that chronically ill patients healthy patients can use in the home to passively monitor their health, and we can flag issues before they become severe disease, we can help chronically ill patients manage that severe disease, and then avoid hospitalizations and ER visits, which are really what cause all the dollars to be spent in American healthcare today. I do think that by making  healthcare less expensive, it would significantly propel the middle class in a direction that would make them live like royalty. And so for me, I think that's what we're building towards. How can we make it so that you can get world class health care in your home  24/7 and on demand? When you want that blood test, you get it.  When you want quick consultation, you get it. When your blood pressure is off, the sensor picks up on that, and flags it. Is this patient at risk for developing hypertension or developing congestive heart failure? How do we help make it so that people can live longer, healthier lives while simultaneously putting wealth back into the American economy? So that's the sort of grand vision. I think there's a lot of steps to get there. And in the process, is a massive company to build in doing so.

 

Pieter Abbeel: Well, I hope that succeeds. I'm kind of curious at the individual level, how do you think about preventative health care? Are there some diseases that come to your mind where they’re just typically detected too late? And it would be so much better if we had early detection?

 

Tanay Tandon: Yeah, definitely. I think there's a whole range of them.  There's both disease itself and then there's also the acute events that lead to hospitalizations for patients. We already have disease. And so, you know, on the front of those acute events. In cancer populations,  hospital utilization is way too high. Often, a patient will become neutropenic, or be at risk of catching an infection, and just won't know. And so step one, I think it's building it so we never miss a neutropenic or event. And the way you do that is you get these sensors into the homes of cancer patients. Similarly, there's really simple diseases as well. When you look at congestive heart failure, you look at long term chronic hypertension and the types of things that occur with heart disease and for cardiac arrest. One of the biggest killers in the country and drains economically on the healthcare system but its so simple to detect so just you know, tracking weight day over day tracking blood pressure. It's so simple but the technology infrastructure doesn't yet exist in order to passively monitor that and pair it with good insights from a physician or a nurse so that's something we're working on. We have 1000s of blood pressure monitors and weighing scales that we ship every month to get into the homes of patients. [unintelligible] It happens but it can definitely be spotted earlier and earlier. The postoperative setting where people might have sepsis, like a severe infection and don’t know about it. Again, it's something where if you sent them home with an Athelas device and you could get a frequent platelet white blood cell hemoglobin count again and again and again, and then spot that spike or spot that trend, and then avoid it from turning into a hospital visit or into death. So those are some of the areas that are low hanging fruit. Then long term, I think there's so much signal in our blood, and there's so much signal in our physiological parameters that realistically, we should be able to spot a whole host of diseases earlier, everything from breast cancer to the risk for developing an autoimmune disease. These are things that physiologically will appear. We're just not measuring that signal. And no one is training models on that signal. No one is actually trying to do interesting, interesting things with that signal to intervene earlier. And so I would say almost all diseases should be detected earlier with technology that we already have.

 

Pieter Abbeel: Well, that's an exciting future ahead. Thanks for working on it.