Skip to content
YC Root AccessYC Root Access

Building the Agentic AI Platform for Hospitals

Bunkerhill Health recently raised $55M to help build a true state-of-the-art AI platform for hospitals. In this episode of Founder Firesides, YC's Ankit Gupta sat down with their co-founder & CEO Nishith Khandwala to discuss how their tools dramatically speed up hospital operations, the cold email that landed them Cleveland Clinic as their first customer, and a future where even the most complicated surgeries are managed end-to-end by agents. https://www.bunkerhillhealth.com Chapters: 00:00 — $55M Series B Announcement 00:52 — What Bunker Hill Health Does 03:50 — Why It Takes Two Years to Onboard One AI Tool 05:09 — The Platform: Knowledge, Reasoning, Action 07:12 — The Innovator's Burnout Problem 09:57 — How Bunker Hill Actually Solves This 13:01 — How Nish Got Into Healthcare AI 17:00 — His Dad's Heart Attack Changed Everything 19:32 — How LLMs Transformed the Opportunity 22:05 — Cold-Emailing Cleveland Clinic 25:18 — Finding the Right Abstraction 28:04 — How Different Are Hospitals From Each Other? 31:12 — LLMs, Tool Use, and Hallucination 34:37 — Measuring Against the Standard of Care 39:27 — Building a Team of 21 43:30 — The Turkey Bone Patient Journey Apply to Y Combinator: https://www.ycombinator.com/apply Work at a startup: https://www.ycombinator.com/jobs

Ankit GuptahostNishith Khandwalaguest
Jul 16, 202647mWatch on YouTube ↗

EVERY SPOKEN WORD

  1. 0:00 – 0:52

    $55M Series B Announcement

    1. AG

      [upbeat music] Today we're really excited to have Nish Khandwala from Bunker Hill Health. Thanks so much for joining us.

    2. NK

      Thanks for having me, Ankit.

    3. AG

      Nish, you have a exciting announcement to share today. Why don't we kick things off with that?

    4. NK

      Yes, absolutely. Uh, we are very, very excited to announce our Series B today. Uh, we have raised $55 million thus far, and this round was led by Vinod Khosla at Khosla Ventures. And joining this round is obviously Y Combinator, um, also Alfred Lin from Sequoia, uh, Optum Ventures, Felicis, and this all will go into building the Agentic AI platform for health systems. Um, we think that this is the way how AI gets adopted in the future, and I'm excited to share more about what we're up to.

    5. AG

      So tell us a little bit about Bunker Hill Health.

    6. NK

      Yes.

    7. AG

      Um,

  2. 0:52 – 3:50

    What Bunker Hill Health Does

    1. AG

      so maybe to begin, what does Bunker Hill Health do today? Like, give us the full span of kind of what the company does.

    2. NK

      Absolutely. We are really obsessed about lowering the cost of iteration in healthcare. I think healthcare has this reputation of it's hard to move the needle in, or it takes a long time to get something done. We are really obsessed about changing that. How can we make the speed of iteration faster? How can we make the cost of iterations less? So what do I mean by that? Um, I'll give you two sides. Let's take the side of an executive at a health system, at a hospital, where this executive is responsible for a- all AI initiatives, and from their point of view, what's going on? In their shoes, they're thinking, "Wow, there's so much potential of AI." I think we are well past that, like, oh, is there potential of AI or not? Like, it is so obvious that there is so much potential to make better patient care happen, drive additional revenue, save cost, drive business objectives. There's so much low-hanging fruit even. But it doesn't seem from their standpoint that there's a viable pathway to actually capturing that potential.

    3. AG

      So, like, they're, they're presumably, you know, they're already using ChatGPT in their personal lives or whatever.

    4. NK

      Of course, yeah.

    5. AG

      Like, they're aware that AI-

    6. NK

      Yes

    7. AG

      ... is amazing, but they don't necessarily see how to incorporate it in the hospital setting.

    8. NK

      Yeah. So again, if you're, let's, let's say I'm the f- that physician, um, or that executive at the health system responsible for AI initiatives, I have people coming to me left and right across the organization. I have the chief of vascular surgery coming to me, saying, "I want an AI tool to track aneurysms so that we can deliver better patient care. Oh, and by the way, it'll also drive better business, so great." I have the chief of radiology coming to me with, uh, with solutions that they want to adopt, the chief of cardiology. I have the revenue cycle team, the quality team. You name it, they're all coming to me with saying, "Hey, we wanna adopt this AI tool." And in that position as an executive, again, this hypothetical scenario where I'm the executive at the health system, uh, responsible for AI, I am not questioning at all whether these AI tools will deliver or have the potential to deliver value, that the value proposition makes sense. It also, like, you can just see the benefit.

    9. AG

      And the teams want it.

    10. NK

      Yes, the teams want it, too. However, this was a bitter lesson that we learned through many years of working in this industry, that the effort that it takes to onboard a new tool far outweighs the benefits. It's like you will spend two years trying to onboard a new tool, and only for it to deliver one specific thing. Imagine if for every app you ever used on your iPhone, you had to buy a separate phone.

    11. AG

      Right.

    12. NK

      Like, how would that feel? Like, it would just be like, eh, can I live without

  3. 3:50 – 5:09

    Why It Takes Two Years to Onboard One AI Tool

    1. NK

      it? Probably.

    2. AG

      And why does it take two years to onboard one of these things? Like, what are those sources of pain-

    3. NK

      So-

    4. AG

      ... in these places?

    5. NK

      It's because the barriers of entry are very high. Um, if the health system saying, is saying, "Yes, we are interested," you need to get through procurement. You have IT security. You have data privacy. You have all the other sort of contracting and procurement-related things that need to happen. You then actually need to get IT resources to implement the solution, and IT resources are shared across different departments. Everybody's sort of fighting for the same resource. Um, you get your day in court at some point, where you finally get IT resources, but then you have to go through ta- change management. Um, there's just so much to be done, where on paper it doesn't seem like a lot. It's like, why can't this be done in two, three months? But when you actually roll it out, it naturally takes a year plus to get that done. So again, from that viewpoint of that physician or that executive at the health system responsible for AI, I am like, "There is no way I can incorporate all of these tools." So it feels almost like I see the potential, but I do not know of a good way to capture it. The insight there is that if you look underneath the hood of all of these sort of different solutions, they look very similar, where they are extracting data from different systems of record.

  4. 5:09 – 7:12

    The Platform: Knowledge, Reasoning, Action

    1. AG

      Presumably EHRs or whatever.

    2. NK

      Exactly, and then putting that through large language models. And as these large language models have gotten better and better, the requirement for fine-tuning or doing any additional machine learning work on top of it is also reduced. And so in that position, in that hypothetical position where I'm the executive, I am thinking, "Why am I going to buy so many different solutions?" It's not so much the economic cost. That's there, but it's not the major factor for lack of adoption. It's more so it's just going to be so much work to do all of this.

    3. AG

      Why go through 10 different procurements-

    4. NK

      Exactly

    5. AG

      ... for something where the two most important pieces are-

    6. NK

      Yeah

    7. AG

      ... kind of similar.

    8. NK

      So that's from the side of that executive at the health system. Now let me tell you the inverse side or the flip side, where, let's say I am a, a chief of metabolic health, and so far, for my patients who were diagnosed with this disease called MASH, it's a type of fatty liver disease, up until recently, I had no treatment to give them. But last August, Wegovy, Ozempic's cousin, got FDA approved for fatty liver disease, for MASH.

    9. AG

      Right.

    10. NK

      So suddenly, I can make so much impact for my patients, and rather than, you know, manually combing through every patient who has come and seen me, not going to be possible, who has come through my department. I mean, it may be possible, but you know, untenable from a labor perspective. And in that case, like the application of these large language models to comb through every patient in my department who has a diagnosis of MASH, not on Wegovy, but eligible for it, seems like a very natural application. Now, I am here thinking, "Hey, I have this great idea for how I can make much... Take much better care of my patients." And I want to do something, but I am getting nos from everybody. I'm getting no from the AI committee, I'm getting no from the IT bandwidth, and it's not because... Like, they're putting up well-intentioned pushback on this. Um, it's just that it doesn't seem like there's

  5. 7:12 – 9:57

    The Innovator's Burnout Problem

    1. NK

      a way.

    2. AG

      Presumably the pushback being they don't want you to just, you know, give your EHR to the OpenAI API or something like that or just-

    3. NK

      It's less so. I, I haven't actually found that... I mean, you need to have the appropriate guardrails, that's for sure. Like any company that is in the space needs to have a business associates agreement-

    4. AG

      Sure

    5. NK

      ... with all the frontier labs so that they can share. I don't think that's, that's really the crux of it. It's more so I'm a, I'm a lead of an initiative. I really want to do this. Heck, I can even build something for it myself, uh, if I have the ability to do so. In my previous life, when I was a researcher at Stanford, we would build all of these kinds of tools all day, day in, day out. But then you would just go through this burnout of you build something or you have an idea for something, it goes nowhere.

    6. AG

      Right.

    7. NK

      Constantly. It's like it's a very vicious cycle. You build something, it goes nowhere, and suddenly you're just expected to show up for your next initiative without any loss of excitement. That's just not gonna work, and it's an energy depleting thing where time and time again, you become less excited about the next initiative. So we want the cost of iteration to go down, the speed of iteration to go higher, which means for that physician executive who's responsible for AI, giving them a way to say yes to a lot of different initiatives, and for that innovator, for that researcher, for that initiative, uh, like the person owning the initiative, for them to actually have an outlet for their innovative ideas. If you do that, I think that the cost of doing something and not working goes down so much, and I think that will result in much faster improvement of outcomes. Healthcare will go from being a laggard industry for adoption to one of the earliest adopters.

    8. AG

      So basically to, to unblock all the stuff that makes it so that people can't actually be experimental, and basically allow people to-

    9. NK

      Yes

    10. AG

      ... run things where they may not be the permanent solution-

    11. NK

      Mm-hmm

    12. AG

      ... but there aren't all these sort of artificial things blocking them from doing them in the first place.

    13. NK

      Yes, exactly.

    14. AG

      Yeah.

    15. NK

      Like I, I want a future in which that, again, that executive from the executive side, it's much easier to say, "Let's just try it out. If it doesn't work, we'll turn it off. Big deal."

    16. AG

      Sure.

    17. NK

      Right? Rather than let's m- we meet on a monthly basis, create a 40-page report.

    18. AG

      Have a task force.

    19. NK

      Exactly, and, and so everybody wants things to improve. Nobody's actively trying to block it. So that's, that's what the future should look like. And from the innovator side, I am hoping that there is a day where there's a researcher or an innova- innovator who is complaining to Bunker Hill about, "Hey, I made this yesterday. Why isn't it already live yet enterprise-wide?"

    20. AG

      Right.

    21. NK

      You know, that almost like lack of awareness at how difficult life was prior to this, that would be amazing. I really want that future.

  6. 9:57 – 13:01

    How Bunker Hill Actually Solves This

    1. AG

      So, so then let, let's talk about the present-

    2. NK

      Yeah

    3. AG

      ... present of Bunker Hill.

    4. NK

      Yes.

    5. AG

      So where does Bunker Hill fit in to actually solve this particular set of problems-

    6. NK

      Yes

    7. AG

      ... you're delineating? Like, what, what's the, um... Yeah, what, what is the, the product offering that makes it so that problem kind of goes away or at least is heavily mitigated right now?

    8. NK

      So we are thinking that there needs to be a platform that has all the integrations that can read in and write out. So think of connectors. You can connect to the EHR, you can connect with other sources of truth, like the ERP of the health system. Uh, you can connect to the financial record, like the source of truth. You can also connect to, like hundreds of other portals that health systems use, whether that's payer portals or registry portals, so on and so forth. And then at the crux of it, obviously you have AI, which can be your favorite large language model. It can also be bespoke sort of vision-based models that are doing more diagnostic things, things like that.

    9. AG

      So you guys have built integrations to all of these things.

    10. NK

      Exactly, and so when we get into a health system, we lead with the idea of a platform. Like we have this concept that a platform should have three pillars: knowledge, reasoning, and action. Knowledge is how we obtain information in. So that's, those are the connectors that are doing the read activity. So common sources would be the EHR, the electronic health record, um, imaging archives, EKG archives, ARPs, things like that, and also the internet. You know, there are a lot of clinical guidelines on there. So that's the knowledge work. That feeds into the reasoning brick, which is where the AI lives, and then AI gets you an answer, which is great. But if we don't take an automated action on top of that, you just have a dashboard that allows you to admire the problem [laughs] as opposed to actually do something around it.

    11. AG

      Yeah. [laughs]

    12. NK

      And so the re- output of the reasoning brick gets fit- fed into the action brick where we can notify patients. We have, you know, auto snail mailed patients.

    13. AG

      Cool. Nice.

    14. NK

      Uh, we auto texted, uh, SMS to patients. Um, we have obviously through the EHR, AI voice calling to some more recent, uh, uh, level. We can message providers through the EHR through automatic fax, if you want to believe that.

    15. AG

      Nice.

    16. NK

      Um, we can write back to-

    17. AG

      That's hilarious. [laughs]

    18. NK

      Exactly. Um, and then we can write back to portals and things like that. So if you have that sort of platform available, and I truly mean platform, like it's not a point solution, it's not a multi-point solution. I mean a true platform that people can build on top. So you have that harness built. That innovators can build on top of. Today, they build with our help, some with their own, like it's a self-serve thing. Tomorrow, I'm thinking it's gonna go much more self-serve as well.

    19. AG

      Cool.

    20. NK

      Yeah.

    21. AG

      That's... Okay, so I have a lot of questions about-

    22. NK

      Yes

    23. AG

      ... your product.

    24. NK

      Yes.

    25. AG

      But maybe before we dive into all of those, I actually wanna rewind a little bit-

    26. NK

      Mm-hmm

    27. AG

      ... to the very early days.

    28. NK

      Yeah.

    29. AG

      You, you mentioned earlier that you were a researcher at Stanford.

    30. NK

      Yes.

  7. 13:01 – 17:00

    How Nish Got Into Healthcare AI

    1. NK

      Yeah, it did not come from the healthcare world.

    2. AG

      Okay.

    3. NK

      Um, came from the AI world, worked in a lab that was trying to make more training efficient, uh, processes for learning. And-

    4. AG

      You were an undergrad at Stanford when you were doing this?

    5. NK

      Undergrad and... Yeah. It was a blurry line, undergrad and master's, but, um, I was a part of a lab. For us, healthcare was just an interesting data set that we could show. Healthcare was specifically one where you could e- even argue that you don't have a lot of explicit labeled data, so we were trying to figure out semi-supervised or unsupervised ways of building models, so healthcare was a great data set for us. And in the process, we ended up collaborating with a lot of different stakeholders within the School of Medicine. I remember we had the chief of preventive cardiology approach us with an idea that we thought was like the best thing since sliced bread. It was a very fun idea. He was like, as a preventive cardiologist, his job is to prevent future heart attacks for patients, so many times he would actually see patients right after they've had their first heart attack, and his job is to prevent the next one. And when this cardiologist would see those patients, he would review their charts, understand what's going on, what led to the first heart attack, and he would find that many times that same patient had come to Stanford Hospital for some entirely unrelated reason, maybe a car accident, maybe lung cancer screening, what have you, gotten a CT scan of the chest where the heart was visible, where the arteries around the heart were visible, and it was evident from that scan that the block- there were blockages, there were plaque buildup in the arteries, but no one did anything about it. The patient fell through care gaps, literally needed a heart, a heart attack before they saw a cardiologist. And so his proposal to our lab was, "Could you build an AI model that could comb through every scan that is being done at Stanford? Basically, you screen every patient coming to Stanford Hospital and see if they're at high risk of heart disease, and if so, bring them to the cardiologist." Coming from the AI world, we thought, "Wow," like, great application for clinical care. Who doesn't want to prevent heart attacks?

    6. AG

      Right.

    7. NK

      And I'm not gonna question-

    8. AG

      Pretty simple computer vision problem too.

    9. NK

      E- exactly too.

    10. AG

      Yeah.

    11. NK

      Like, I... At no point was I doubting whether or not this model was possible to build. Um, and then too-

    12. AG

      E- especially at that time, right? This is like 2018 or so?

    13. NK

      2017, 2018, yes, yeah.

    14. AG

      Right, so this is like right at, like computer vision kind of works.

    15. NK

      Exactly.

    16. AG

      We know how to do this kind of thing.

    17. NK

      Yes, exactly.

    18. AG

      If we have enough labels, which they do, seems like a straightforward problem.

    19. NK

      Exactly, and so... And the s- the other important thing was that we also saw an immediate economic case for this as well. This wasn't yet another like, oh yeah, this makes sense in an idealistic world, but hey, because this is not driving economic impact anywhere, it's not gonna get adopted. Um, we live in a mostly a fee-for-service world where health systems get paid for providing services, so if this tool gets built and gets adopted, this means that the cardiology department at a hospital will see more patients, and seeing more patients means more revenue. And so we thought, "Wow," like unicorn use case, right? Good for patients, good for revenue as well, and so we thought, "Slam dunk." We built it. It worked really well, and I could not tell you how much I was excited to, like actually finally get this deployed. And then it was-

    20. AG

      And this, this is all still as a student?

    21. NK

      All as a student.

    22. AG

      Not, not, not a startup or anything like that.

    23. NK

      Yes, exactly.

    24. AG

      Cool.

    25. NK

      And, and it was the simplest thing we were trying to do, which is I was at Stanford build- with a Stanford-built model trying to deploy it at the Stanford Hospital. Like, there was no externality, nothing. Like, data was gonna stay at Stanford, nothing. Like, there were no logical reasons for this to not get implemented. But we realized that it... Like, I heard the words, "It's someone else's job," and I was like, "Who is that someone else? Please let me know." [laughs]

    26. AG

      [laughs]

    27. NK

      Uh, and so for many years we became that someone else, uh, and just trying to figure out what it takes to deploy a model. We started by doing it at Stanford, and then we wanted to do it at other hospitals too, like UCSF, um, uh, uh, MedStar in Washington, DC, and several other health systems. And there was just so much interest from everybody to do this, but we found that it was so darn difficult to get anything done.

  8. 17:00 – 19:32

    His Dad's Heart Attack Changed Everything

    1. NK

      Um, and the worst part of all of this was that if we wanted to do this for the next model, we would have to reinvest all of that again.

    2. AG

      Yeah, interesting. Like, it's not even like you get a head start because you've done-

    3. NK

      Nothing

    4. AG

      ... a ton of stuff.

    5. NK

      Nothing.

    6. AG

      You kind of start from scratch every time.

    7. NK

      Exactly, and so we thought, "Man, this is not gonna go anywhere," and it leads to that burnout. We saw a lot of our lab members leave over time. It's just like, why don't we... And I was ready to sell my soul, go optimize that somewhere as well. [laughs]

    8. AG

      [laughs]

    9. NK

      Um, because again, I didn't come from healthcare.

    10. AG

      Yeah.

    11. NK

      So I was like, "Cool problem," but like-

    12. AG

      Who wants to deal with this? Yeah.

    13. NK

      Exactly, um, but then my dad got a heart attack, had a heart attack, and he, he's fine now, thankfully.

    14. AG

      That's very good.

    15. NK

      Um, that was five years ago now. Um, and the cardiologist calls me, so they live in Abu Dhabi near Dubai, and the cardiologist calls me and it's like, "Oh, we did this CT scan on your dad and we found this plaque in his coronary arteries. Here's what it means." And I'm like-

    16. AG

      Like, you gotta be kidding me.

    17. NK

      Exactly. It's like, are you kidding me? Like, I s- I had spent many years of my life working towards deploying this tool, and I had practically given up on doing that, but then this happens, and so this professional frustration married with the personal frustration, like we became students of this industry. Uh, my co-founder friend at that time was also facing a similar issue where he built something and it didn't seem like there was a pathway to actually get it deployed, and so I'm very glad that he was also having that problem. Because had it not been for that, a very logical next step for us would have been to just create a company for that one algorithm, for that one use case. But then we-

    18. AG

      Oh, interesting, but he was also like a computer scientist-

    19. NK

      Yes, exactly

    20. AG

      ... interested in healthcare or something.

    21. NK

      Yes.

    22. AG

      But he was blocked by some other algorithm being-

    23. NK

      Yes

    24. AG

      ... deployed to the hospital. Okay, interesting.

    25. NK

      Yeah, exactly.

    26. AG

      Cool.

    27. NK

      So he was working on AI for robotic surgery, and it was like, oh, y- you need like these models to be validated, deployed, FDA cleared, so on and so forth. There's no way I can see an end to this. And so I'm very glad because it allowed us to take a step back and see that pattern that's like, "Hey, we are not unique in having this problem." And that problem got super exacerbated by LLMs because now many algorithms just became prompts.

    28. AG

      Right.

    29. NK

      So we no longer even needed to solve the problem necessarily of you build something at one hospital, how do you get it out everywhere else? It was more so people at the hospital itself are going to be able to build these models very quickly, uh, or these use cases, how do you actually get them deployed? So that's where,

  9. 19:32 – 22:05

    How LLMs Transformed the Opportunity

    1. NK

      uh-

    2. AG

      It's interesting 'cause it both exacerbated the problem-

    3. NK

      Yes

    4. AG

      ... because there's way more things you could-

    5. NK

      Mm-hmm

    6. AG

      ... now do with these much easier to program algorithms-

    7. NK

      Yes

    8. AG

      ... so just prompts, but it also, I guess, simplifies the deployment to some degree because, well, you just need to pick some LLMs, and then you can solve a huge number of problems as opposed to before you needed some like, you know, bespoke deep learning pipeline-

    9. NK

      Yeah

    10. AG

      ... for each different use case, and then for each of them, figure out how they're gonna deploy it through a single infrastructure, and they might need very different amounts of compute and whatever. Like, all of that basically goes away.

    11. NK

      Exactly, and there's another aspect, too. So when you are choosing, in the, in the pre-LLM world, when you were choosing which problem to work on, naturally you would gravitate more towards the clinical problems because those are cooler, um, and every use case required bespoke algorithm building, especially ones that required you reading over text. And so if you look at like the Maslow's hierarchy, there are some problems at the base level that a health system faces, uh, the cl- care delivery system faces, and then there are some like at the very top that are enlightened use cases.

    12. AG

      [laughs]

    13. NK

      And everyone was just going after the enlightened use case-

    14. AG

      Right

    15. NK

      ... because where else, where, where else would researchers spend time on?

    16. AG

      Naturally.

    17. NK

      But with LLMs, we now actually had a way to offer solutions for every problem, every type of problem within the Maslow's hierarchy, um, such that a health system could say, "Hey, we want to invest in this type of platform. Yes, we will solve the hair on fire, the important and urgent problems for today," but it also gives a way for everybody else for their enlightened problems to be solved as well, and I find that really cool.

    18. AG

      Totally.

    19. NK

      Ironically, by us offering a platform to do the non-cool stuff, more cool stuff gets adopted as well.

    20. AG

      Right. You're offering the very bottom rung-

    21. NK

      Yes, exactly

    22. AG

      ... of Maslow's hierarchy.

    23. NK

      Yes.

    24. AG

      Interesting.

    25. NK

      Yes, exactly. Um, and so the LLMs have just been like a huge boon to us. Uh, yeah.

    26. AG

      Yeah. So when you, when you then got started and you realized that, okay, there was this opportunity to solve this kind of fundamental thing that-

    27. NK

      Yeah

    28. AG

      ... prevents all of the higher level things from occurring, e- even so, I could imagine you at the beginning are like, "Okay, how do I actually like make the first version of this-

    29. NK

      Yeah

    30. AG

      ... that a hospital is willing to trust the two of us in?"

  10. 22:05 – 25:18

    Cold-Emailing Cleveland Clinic

    1. NK

      So I'll go back to that C team base model for heart disease detection. We didn't know any better. We were just like, "Let's figure out what it looks like with this one model, and then we'll generalize it to everything else from there." Um, very luckily for us, um, a cold email to a Cleveland Clinic physician resulted in Cleveland Clinic being our first customer.

    2. AG

      That's awesome. [laughs]

    3. NK

      I don't think-

    4. AG

      That's epic

    5. NK

      ... Cleveland Clinic actually knows [laughs] that we were our first customer.

    6. AG

      Well, they do now, I guess. [laughs]

    7. NK

      But I'm sure, uh, I'm sure working with us they probably found out.

    8. AG

      [laughs]

    9. NK

      But you really need someone like who's just an early believer, where it's like, "Hey, look, this is new. I can't tell you we have a playbook for this because nobody has done this." So we took that model-

    10. AG

      So you found an early adopter.

    11. NK

      Correct. We... And again, you know, it's like luck, needle in a haystack because it was a cold email, right? Um-

    12. AG

      Yeah, but like you kind of engineered that luck for yourself by sending the email.

    13. NK

      [laughs] Of course, yes.

    14. AG

      Like a lot of people would just be like, "Oh-

    15. NK

      Yeah

    16. AG

      ... this doctor's never gonna reply to me, so I'm not even gonna bother."

    17. NK

      Right. Um, but it really was we took that first model, and we were like, "Let's just figure out what deployment looks like at that large of a scale." So Cleveland Clinic is a ginormous system, and when we ran a retrospective study to see how many patients we would find, we saw that it's going to be in the thousands, and when you see a number that large, we were like, you know, we would get laughed out if we would go to go back to them and say, "Here's a list of 9,000 patients. Here's our invoice. Thank you so much."

    18. AG

      Right.

    19. NK

      And just expect them to magically conjure up like 20 to 30 nurses to navigate then follow up with those patients. Um, that's where sort of the genesis of the action bit came about. And when we proposed like, oh, you know, we are gonna receive the scan, run the model, uh, run LLMs on the patient's record, see which patients need to be followed up, and then automatically notify them, then with them and with other partners, early adopters like UTMB, we quickly figured out that this was the right abstraction to work on, that knowledge, reasoning, and action mapped to so many different problems. It was like, oh, we want to identify patients in our population that are currently uninsured but should be on Medicaid but are not.

    20. AG

      Right.

    21. NK

      So could we run an AI agent to comb through every patient knowledge, reason whether the patient is currently uninsured plus eligible for Medicaid? And then let connect, like, the action brick would be connecting those patients to social service, social workers, things like that. That's another very different use case, same abstraction. You take something much more common, like revenue cycle. It's like, oh, the hospital needs to tell insurance companies what they have done so that they can get appropriately billed for. Knowledge, you pull in patients' information, you see what's being done, you use AI to figure out what billing codes need to be applied, and you take the action of actually submitting those codes. As you can see, like, you can have clinical problems, administrative problems, clinical ops, the rev cycle, prior auth all fits that paradigm, and that's when things just clicked for us, which is like we have found the right abstraction to work

  11. 25:18 – 28:04

    Finding the Right Abstraction

    1. NK

      off on this.

    2. AG

      Yeah, it's interesting how you describe that. Um, it's a very computer scientist way [laughs] of describing the problem, right?

    3. NK

      Yeah.

    4. AG

      You're, you're describing it in terms of, like, you know, an object model essentially.

    5. NK

      Yeah, exactly.

    6. AG

      Of like, you know, how, how information transfers between these. And, and I guess, like, related to that, you know, it may not be as obvious to the outside what types of hard engineering problems it takes to actually service this.

    7. NK

      Yeah.

    8. AG

      Like, you s- you've so far described a lot of, I guess, kind of organizational problems-

    9. NK

      Yes

    10. AG

      ... of how you get the right people to agree and find the right stakeholders and then connect the right systems. Maybe even just, like, diving into the engineering a little bit-

    11. NK

      Yeah

    12. AG

      ... like what are some of the hard engineering problems that come up when you actually do that and deploy to, like, Cleveland Clinic, who is a enormous enterprise customer-

    13. NK

      Yeah

    14. AG

      ... to work with?

    15. NK

      I- it's more about what is the end goal, and then the technical problems sort of follow that. So we want to, again, decrease the cost of iteration, increase the speed of iteration. Now, what, what does that mean? If we have to build a platform where we, a hospital adopts the platform, and then we have engineers who then need to, like, forward deploy people who then need to send out a product brief to our engineering team, and the engineering team then builds that and then delivers, then we'll be one of those companies that promises a lot but not delivers very-

    16. AG

      That sounds kind of slow.

    17. NK

      Yes, it's very slow

    18. AG

      ... that sounds like a slow iteration of the process.

    19. NK

      So we, the, the technical challenge here was to land on the right primitives, the right set of abstractions. So today, our forward deploy team, when they work with different stakeholders within the health system, they do not send anything back to the engineering team saying, "Build this for me." The product needs to be built in a way where the forward deploy team can be non-technical people, obviously very smart, um, incredible critical thinkers, but they can just use the platform itself to service the end user, to service that stakeholder. And so c- that is a very difficult problem because it is such a complex industry where you have to land on the right primitives. Otherwise, you're not going to really deliver on this, which is why, you know, people are usually shocked when we tell them that when we work with the health system, we work with them on 15, 20 different use cases. You're not looking at one, two, three use case. You're looking at 15 plus use cases, and that does not happen without that right level of-

    20. AG

      Like, you have to have the set of primitives such that you can-

    21. NK

      Yes

    22. AG

      ... you actually work on 15 or 20 different use cases or something.

    23. NK

      Exactly, without having to do engineering, so.

    24. AG

      How much are hospitals even similar to each other when it comes to the systems they use? I mean, I'm sure there's some system that-

    25. NK

      Yeah

    26. AG

      ... basically all of them use. But I think of these as, you know, a lot of these hospitals have built up over such a long period of time that they might have some, like, uh, their, their technology stacks might be like entirely different from one another. H- how do you actually navigate that?

  12. 28:04 – 31:12

    How Different Are Hospitals From Each Other?

    1. NK

      So thankfully, the technology is, the technology stack is pretty similar hospital to hospital. Um, for whatever reason, Naomi can go into the history behind this, but all the HRs adopt the same standards, FHIR, HL7. The, it's, but it's not like, without LLMs, nothing could still happen because these are like PNGs or JPEGs-

    2. AG

      Sure [laughs]

    3. NK

      ... but they're the same format, but in the inside of it could be whatever, right? In the same way, HL7 and FHIR standards are adopted universally across health systems, or at least that's been our experience when we work with health systems that are over a billion dollars of revenue. Um, and in that, so, you know, it's not a technological problem per se. Imaging is also very standardized, so on and so forth. So that's not the issue. However, every hospital is different from each other in the sense that their processes look different. Everybody does the same thing in a slightly different way.

    4. AG

      I see.

    5. NK

      So if you build a more traditional software where you expect the health system to undergo change management and adapt to your sort of, uh, assumed workflow-

    6. AG

      They're gonna be like, "Take a hike."

    7. NK

      Exactly.

    8. AG

      [laughs]

    9. NK

      So that's another reason why the primitives matter a lot, so that you can quickly just shadow what they're doing, understand what they're doing, and rearrange the primitives to solve for their exact workflow.

    10. AG

      Could you give me an example of that?

    11. NK

      Yeah, yeah.

    12. AG

      Where, where, like, two different hospitals are doing kind of, you know, at, at face value-

    13. NK

      Yeah

    14. AG

      ... are doing kind of the same thing, like, I don't know, uh, a particular treatment or something like that.

    15. NK

      Yeah, yeah.

    16. AG

      But the processes they're running are sufficiently different, where, like, a very static, non-customizable set of primitives-

    17. NK

      Yeah

    18. AG

      ... wouldn't be able to satisfy both of them.

    19. NK

      So even problems that look similar, like there's this problem where patient could go in, get a scan for whatever reason, and they might have an incidental lung nodule, um, like an X-ray or a CT. And I was shocked as to how different two different hospitals could look in terms of what they think. It's like, "Oh, we want to risk stratify based on the Fleischner guidelines. We want to risk, risk stratify based on a Brock score." And it's just like, okay, well, you can't have the Fleischner people adopt the Brock score or the Brock s- otherwise, you need to change their way of thinking as opposed to build a software to service their needs. That's a simple use case. Uh, we have had health systems where, um, they are trying to do a process manually before AI existed, things like that, and they created a really sophisticated process behind it. They, it's like a team of 20, 30 people that is doing something bespoke and very manually. And because AI was not a thing of the past, they came up with processes that are very sophisticated for their specific purpose. Um, and so we could not go to them and say Hey, just put everything that you've already worked on in the trash. Let's start from scratch. Here's the entirely new process. That would've taken months, if not years, to get adoption for. Uh, but if we are like, "Oh, this is not that different, but it's different enough where, you know, the existing model, existing, uh, software could not work," then here we can go and quickly arrange the primitives, the bricks that we call. Like, our platform is called Carebricks.

    20. AG

      Nice.

    21. NK

      Bunch of bricks put together. So we'll rearrange the bricks to solve that problem for

  13. 31:12 – 34:37

    LLMs, Tool Use, and Hallucination

    1. NK

      them.

    2. AG

      Interesting, yeah.

    3. NK

      Yeah.

    4. AG

      And, um, uh, you know, it comes to mind, you know, when you started this, you started in 2019.

    5. NK

      Mm-hmm.

    6. AG

      At the time, LLMs were not out. I imagine when that moment occurred of ChatGPT coming out, or at least, uh, maybe for you, you were looking at the research even sooner, so, you know, experimenting with GPT-2 and whatnot, I'm sure you saw that as a disruptive force to whatever you were doing before. And so I'm curious if... Are, are there thoughts, uh, do you, do you have any thoughts to share around in that moment when that technology came out and you could now finally take, uh, pretty unstructured data, or at least, uh, inconsistently organized data, and do useful things with it? How did you guys react to that and kinda reposition yourself to-

    7. NK

      Yeah

    8. AG

      ... deal with that well?

    9. NK

      Um-

    10. AG

      That was like two years into your company.

    11. NK

      Yeah, yeah.

    12. AG

      Or two or three years into your company, right?

    13. NK

      Three years into the company. Um, we were still students of the industry, and still are, but we, we were just like assume the status quo, what does it take to deploy models, things like that. So I wouldn't say we were, like, commercial, but what we saw as a pattern was that because text data was so difficult... Like, NLP, people forget that NLP was much harder than vision back in 2017, 2018.

    14. AG

      Oh, yeah, I remember. Yeah.

    15. NK

      It's like everybody thought that vision would get solved first, then language would get solved.

    16. AG

      I was working in a NLP lab for that reason. We were like, "Oh, cool, seems like vision is sort of figured out by now-"

    17. NK

      Exactly, right?

    18. AG

      "... so let's work on NLP." Yeah.

    19. NK

      And, and so we had a lot of vision models come in early days because that's what worked well. And also, like, if you think about pathology slides, radiology slides, radiology scans, things like that, like, yes, there are differences across health systems because of patient, um, sort of demographic differences, prevalence, disease prevalence, different things like that, but ultimately when... if I ask you to think of a chest X-ray, I also think of a chest X-ray. We'll probably conjure up very similar, uh, images. But text, that's just completely unstructured. So we always thought that unstructured text would be the last, the final boss. [laughs] But so when LLMs were becoming a thing, uh, BERT, you know, we had then GPT-2 and then 3, we're like, "You know what? Like, maybe text goes first, actually." [laughs]

    20. AG

      [laughs] Oh, interesting.

    21. NK

      Um, and then I think a part of this is just, you know, we, we... You have, uh, YC teaches you th- this well, where you don't need to try to understand whether you've hit product market fit or not. You will feel it. And that was a part of it, too, where once we actually had the primitives in place, we had people coming organically to us from the health system saying, "We wanna use the platform for this. We wanna use the platform for that. We wanna use the platform for this." So we didn't... I don't even remember having a conscious conversation of, like, "Hey, we think text is becoming a thing." Like-

    22. AG

      It's very obvious

    23. NK

      ... it, it's just like, it's just coming at us as opposed to us trying to be ahead of the game and trying to, you know, uh, predict what the future was gonna look like. We were like, "The future is here." Uh, and so that, that, that was, um, that was what happened there.

    24. AG

      You know, one of the sources of skepticism you hear about people using LLMs in hospital or medical setting, you hear... I heard this much more, let's say, one and a half or two years ago, less so now, was around, you know, like hallucination-

    25. NK

      Yes

    26. AG

      ... or, like, model accuracy or something to that effect. I'm curious... I, I'm sure when you were getting started this was, like, a very serious problem to think about.

    27. NK

      Yeah.

    28. AG

      I'm curious what, how you thought about the, um, quality of the models over the last-

    29. NK

      Yeah

    30. AG

      ... uh, especially over the last year, but over the last two really, and how that has changed what matters for you guys to focus on as the models have gotten better over the last few years.

  14. 34:37 – 39:27

    Measuring Against the Standard of Care

    1. NK

      The, yes, the LLMs, the base models have become a lot better, but really what has been a much more impactful thing is the tool use.

    2. AG

      Mm.

    3. NK

      You can now browse the internet. You don't need to condense all guidelines and expect the models to just know it. You don't need to condense all payer policies into one sort of, you know, provide that as context to the model. It can search those by itself, learn it on the fly. So tool use has been, like, it has moved the needle a lot more in terms of reducing hallucinations than the just inherent the base model getting better. But obviously that-

    4. AG

      And so that's a relatively recent thing.

    5. NK

      Yes.

    6. AG

      It's really only in the last six months that I would say tool use got very reliable.

    7. NK

      Yes. Um, and then we are seeing more and more use cases just become now possible because of that. Um, and also on that, you know, there are some very fundamental things that we can do in terms of how we architect the system. Let's not have the model spit out anything without a citation back into the s- into the system of record. So if it's saying that the patient has disease X from, as it inferred from the medical record, let's show it back. Um, in fact, there was a, a fun... So there's this thing that hospitals do, which is fill out registries. Think of it as databases, uh, and they do it for the American College of Cardiology, uh, American College of, uh, American Cancer Society, things like that. And there's a particular question which I remember. It's like, "Is the patient active on e-cigarette use?" And the LLM looked up a random physician note from the plastic surgery department where it says, "Patient also vapes nicotine at home," and it sort of inferred that and-

    8. AG

      Wow.

    9. NK

      So now if there was no link back to the EHR, whoever is reviewing those results would have to basically play the game of, like, a needle in haystack, and that would have been an incredibly difficult way to see adoption happen, but this way because you could just link it back. So yeah, tool use and just citations back to the EHR or to the system of record.

    10. AG

      And that example's also cool 'cause it also feels like the type of thing that a human may not have caught 'cause it-

    11. NK

      Mm-hmm

    12. AG

      ... 'cause it's like in a niche, you know-

    13. NK

      Correct

    14. AG

      ... weird one, that that's a huge advantage of having these systems that are like, you know, they basically are just like, "Yep, sir, I'm just gonna go read every single document [laughs] and just, like-"

    15. NK

      There's no way

    16. AG

      "... read everything and give you the full detail."

    17. NK

      It is unreasonable to expect humans to do that. Um, and so yeah, you can just have it do more work than what you would expect a human. Which actually is a great segue into my second, uh, school of thought, which is since we were very academic students of this industry, of like let's see what the value proposition is, how do you validate models, things like that, I have been a strong believer that the performance comparison should not be 100%. It should always be against the standard of care.

    18. AG

      Hmm.

    19. NK

      Now, what's difficult is that we, for most things, we don't even know where the standard of care is.

    20. AG

      Right, we don't really measure-

    21. NK

      Yes, exactly

    22. AG

      ... accuracy or whatever-

    23. NK

      Yeah

    24. AG

      ... of the person.

    25. NK

      But that has been a big push that we are trying to make, which is always try to evaluate these things in terms of how they compare to the standard of care. So there are some use cases for which that bar is really low. If you don't have an existing process because human labor is too expensive and, and too scarce, so you are just not doing something. So now, with the sheer fact of doing something, in most cases, the bar is so low to clear because the, the, the alternative is doing nothing.

    26. AG

      Right. [laughs]

    27. NK

      Uh, obviously, if there's an existing process for something, then you need to show it against, uh, the standard of care. So that's been a huge ad- like something that we advocate a lot for, which is let's make it a comparison against the standard of care, which we can measure based on historical. Now, again, you know, we can just do historical analysis very quickly and get you what was happening before and after.

    28. AG

      And, and how do hospitals think about that as standard? Like, now as you've built these relationships-

    29. NK

      Yeah

    30. AG

      ... with, uh, a number of, uh, quite a few hospitals-

  15. 39:27 – 43:30

    Building a Team of 21

    1. AG

      Yeah. Okay, why don't we change gears slightly-

    2. NK

      Yeah

    3. AG

      ... and talk a little bit about your team.

    4. NK

      Yes.

    5. AG

      I'm, I'm curious, especially in this current technological moment we're in, um, founders are really rethinking or, in some cases, have, have been thinking, uh, about how to architect a team that can use all the incredible technology that's being built really effectively, and you already described some of this earlier. You mentioned you have a product team and sort of forward deployed folks-

    6. NK

      Yeah

    7. AG

      ... who go into the, the hospitals. Um, how have, I guess, like, changes in technology, especially around coding agents, affected how you think about what your team should look like and how lean it should be, and also what everyone's job should even be in, like, building this product and getting it to customers?

    8. NK

      Yeah. It's both team and product. Um, I think we always, when we generally think about how AI changes the shape of teams, we think about it in terms of efficiency, but I think there's also something to be said that the first derivative of the product should also look very different as well. Uh, with all of that said, like, this comes back to our engineering team only works on the, the platform level. No one has ever written... I mean, nobody writes that much code anyways.

    9. AG

      [laughs]

    10. NK

      But nobody is even, uh, like no engineer is working on a specific use case. We only work at the level of abstractions.

    11. AG

      Those primitives.

    12. NK

      It's very important, and there should be no compromise on that. And then the forward deployed team is the one where they can serve multiple different health systems over a dozen of use cases per customer, uh, and really drive the product's adoption across the health system. And so for that reason, like, our engineering team will remain lean, um, and I, I just ma- I just can't see the 10,000-person team anymore.

    13. AG

      Yeah.

    14. NK

      It's just like, why? And, and for what it's worth, I really, really love working in a smaller team.

    15. AG

      Yeah.

    16. NK

      Uh, like we are 21 people today, and we have more customers [laughs] than number of people actually at this point.

    17. AG

      That's awesome. [laughs]

    18. NK

      Uh, which is, which is great, but, um, when you have such a small team, you just have fewer problems. Like, none of our engineers need, like, a product brief of what's going on in the field. The, we, we, uh, every engineer is kind of like a product engineer to some extent. Like, they understand what's important for the customer, and they kind of work backwards from that. Um, so every engineer is much more, like they see what the customer is able to use the product for and see where to go from there. Um, even on go-to-market side of things, whether that's marketing, obviously, you know, there's, there's, the applications of AI are so obvious, but even on more on the fields go-to-market side of things, it's because we can now... One person can do the work of so many different people. It's just amazing to see that happen. Um, and so I want to keep our team. We are still growing. Like, we're gonna hire, so we're gonna go from 20 to a modest 35 by the end [laughs] of the year.

    19. AG

      That's solid.

    20. NK

      Not like 20 to like 100.

    21. AG

      Totally.

    22. NK

      Those days, I feel they're gone, or maybe we're doing something wrong. [laughs] I don't know.

    23. AG

      Probably your token spend will go up faster-

    24. NK

      Oh, yes. Absolutely

    25. AG

      ... than your people spend. Yeah.

    26. NK

      Yes. Absolutely. But, uh, yeah, we think of teams as we add, um... It's very systems oriented, systems thinking again. It's like every stakeholder at the hospital, whether that's the decision maker, whether that's the end users, whether that's the IT team, whether that's procurement, they all are mapped to one- Individual from Bunker Hill

    27. AG

      Oh, cool

    28. NK

      Um, and similarly, again, and no engineer works on anything use case specific, only at the platform level. So because of things like that, I feel like we have a very functioning small team, and I wanna keep it that way. [laughs]

    29. AG

      Totally.

    30. NK

      Uh, it sounds very fun. It, it's a very fun time to be working on something like that.

  16. 43:30 – 47:49

    The Turkey Bone Patient Journey

    1. AG

      Maybe as a final question, I'm curious, you know, in thinking from the perspective of a patient, you know, we've talked-

    2. NK

      Yeah

    3. AG

      ... so far quite a lot from the perspective of hospitals and, um-

    4. NK

      Innovators

    5. AG

      ... uh, innovators in hospitals, and now I'm curious from the perspective of a patient, what does your company mean for a patient interacting with the health system, not necessarily just today, but let's say over the next decade?

    6. NK

      Yeah.

    7. AG

      Like, how should I expect to experience something different, if anything, from my side as a patient?

    8. NK

      Can I walk you through a patient journey?

    9. AG

      Yeah, sure.

    10. NK

      This is a real story.

    11. AG

      This is the, a current patient journey.

    12. NK

      Yes.

    13. AG

      Okay, cool.

    14. NK

      Yeah, yeah. We had a patient last Thanksgiving who had a turkey bone stuck in their throat, and so they came to the emergency room of this partner hospital that we work with, uh, obviously for the turkey bone, and everybody at the hospital was treating the patient for that turkey bone, obviously. Um, but they get a CT scan, and the Stanford model that I described earlier, it works on that scan behind the scenes. Uh, the patient gets discharged after the turkey bone, but then the patient gets an automated letter from us on behalf of the hospital saying, "Hey, you came in. You got the CT scan. We found this, uh, blockages in your arteries. We recommend that you follow up with a cardiologist, especially because given your records, we haven't seen any indication that you're being treated for it right now." The patient comes back to the hospital, now then sees a primary care provider. The primary care provider says, "You need to see a cardiologist," and the patient gets added to the referral queue. Now, this referral queue has thousands of patients, and it's first in, first out in a conventional setting. Can you imagine?

    15. AG

      Right. Oh, God.

    16. NK

      First in, first out.

    17. AG

      Seems suboptimal, yeah.

    18. NK

      Everybody is stat.

    19. AG

      Sure.

    20. NK

      Right? Like, everybody is an urgent referral. So you just have this ginormous queue, and instead of doing first in, first out, there's an AI agent that is combing through every patient in the referral queue, applying an urgency score, and then changing the prioritization order. Which patients can be treated by virtual care? Which patients need to be treated in a brick-and-mortar setting? Which m- physicians should they be matched with because of their subspecialties? So then that patient gets a scheduling done, also through an AI agent, and then finally sees a cardiologist much earlier than they would have seen in the f- first in, first out referral queue. They see the cardiologist. The cardiologist understands the patient's history, and turns out that this turkey bone patient was actually experiencing chest pain but had dismissed it as heartburn. So this patient was symptomatic, and very likely that they would have had a heart attack in the next couple of months or years. And so then the patient undergoes further testing, and as the patient is getting tested, the, the health system needs to ask the insurance company if they're gonna cover the cost. You have an AI agent helping with that as well. Then the patient finally gets, uh, the procedure, so the, in this case, they got a triple vessel bypass surgery, where it's an open heart surgery, and you... It's a very cool surgery, actually. Uh, very cookie cutter these days, but I still find it, like, whoever thought of it, amazing.

    21. AG

      Crazy, yeah.

    22. NK

      Yeah. So the patient gets the surgery, then the follow-up also happens automatic. So across the journey, you know, like whether... And then also for billing, the health system uses an AI agent to then bill the insurance pro- companies. So throughout the journey, whether that's diagnosis, patient sort of like through the continuum of care, all the way from diagnosis to referral patterns, to actually advising on the treatment, to getting in front of the right physicians, to, um, actually then making sure that the hospital gets paid for it, so on and so forth, like, the, for the patient, you can expect a lot more of a connected experience, a lot more, like, as if there's somebody always looking out for you. And hopefully, as AI continues to deliver, there'll be more and more applications of AI that go towards those enlightened use cases-

    23. AG

      Yeah

    24. NK

      ... of, like, earlier diagnosis, better treatment selection, enrollment in clinical trials, helping matching patients to the right therapy, and this can all happen through one system of action, I believe. Information is coming in, you're running AI, and you're taking automated actions.

    25. AG

      That's a pretty inspiring-

    26. NK

      Yeah

    27. AG

      ... present and future.

    28. NK

      Yes.

    29. AG

      Thanks so much for joining us.

    30. NK

      Yes, thanks for having me.

Episode duration: 47:50

Install uListen for AI-powered chat & search across the full episode — Get Full Transcript

Transcript of episode 3lQQadC6vKg

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.

iOSAndroidClaudeChrome