Name: Claude Code in Healthcare: How Physicians are Building with AI
Uploaded: 2026-04-23T18:32:39.883Z
Duration: 1 h 1 min 56 s
Description: Claude Code in Healthcare: How Physicians are Building with AI

Transcript for "Claude Code in Healthcare: How Physicians are Building with AI": We're live. Hi, everyone. I'm I'm guessing I'm supposed to introduce every the the our speakers here. I'm here with with Graham Walker and Michal Neshetko, and, we're gonna talk about, Claude Code in medicine and and and what changes are happening there. I'm sure I have a script somewhere that I'm supposed to be reading off of that I'm not. Oh, there it is. This is very helpful. Okay. I got you, Daisy. So awesome. Yeah. So, I'm mostly gonna shut up in a second, but, I think I oh, there we go. We are recording the session. Questions can be submitted through the question widget over on the side. Also, if you're you're on your computer, you should be able to zoom in to the slides, I believe. Maybe not. But, anyway, yeah. Drop questions in the in the q and a. Give us feedback. There's a survey widget on the side too. Okay. Let me introduce everyone here. I'm doctor Daisy Holman. I am on the Claude Code team here at Anthropic, and I'm I'm here with, Graham Walker and Michal. Do you two wanna introduce yourselves? Sure. Thanks, Daisy. Hi, everybody. I'm so excited to be here. I think there's thousands of people. What an opportunity. I'm Graham Walker. I'm an emergency physician at Kaiser Permanente here in San Francisco. I'm a clinical informaticist and a a founder. I created MD Calc about twenty years ago. It's our twentieth anniversary to help physicians take better care of patients with evidence based medicine and decision support. I am really excited to talk with you guys today about how I've been using Claude Code to just get people more excited and engaged in the clinician community about building and creating because I think we're the ones that we should be building and creating for medicine. Yeah. Awesome. And, my name is Mihau. I am, what you would call the interventional cardiologist. So, basically, basically, I put stents in people. I treat heart attacks, coronary artery disease. I'm originally from Poland, but for the past eight years, I, I've been in Belgium where I practice medicine. And along my medical career, I've also called it a lot. And, last February, I had this crazy story. I went to San Francisco, and I ended up winning, Anthropix hackathon. I ended up third out of 13,000 people with PostVisit AI. It's a cool project that I'm going to share with you today. And, the events have precipitated so far that, my life turned around, and I'm basically moving to San Francisco right now to build AI. So, I'm going also to show you some best practices, some of the systems that, I've created through, last years and how Claude Code accelerated creation of those. Because what's the guys at that Claude Code have have done for the past months, it's just nothing sort of amazing. You guys are shipping like crazy. So, well, back to you, Daisy. I am so excited about this. This is just this is the kind of stuff that gets me out of the bed in the morning, and I I I genuinely am so honored to have, real life doctors talking about the difference you're making in the world with the product that I get the privilege to work on. So, yeah, we're gonna talk very briefly, about, AI's role in clinical care, and then we have a couple demos. I'm gonna mostly get off stage and not talk too much. But we have demos from both of the doctors here, and then we're gonna have, live q and a with the team afterwards. I okay. So, yeah, I mean, AI progress is accelerating. I think that's, generally an accepted thing here, but I think the thing we want to talk about here is how does this impact the way we make people's lives better. I think that I often get jealous of, doctors getting to interact so directly with making the world a better place, with making people's lives better. And, I want to be a part of figuring out what role AI has in that. If any, I'm happy to say that maybe, you know, if we decide that patients need human touch a lot more than we're, thinking about, then that's that is a a happy solution, but I wanna have the conversation. And, yeah, I mean, AI has been used in in various forms, over the years, to, with EKGs and with breast cancer screenings in a more, detective, role. And now we're really starting to see what can we do with really generative, applications of AI, to really meet patients' needs. And and both of our speakers here have excellent examples of that. So, there there definitely is a lot of, at anthropic, we have this this this phrase we've called the the race to the top. Right? Which is, like, how can we do the better thing for society in a competitive way? And, you know, the impact we wanna have medicine on medicine falls right right into this. And so, like, safety and privacy are so important in medicine, obviously. Building trust with your patients, we definitely wanna talk about how these, tools can enable you to build trust with your patients and not detach you from your patients in any way. And and also just performance, like, catching things earlier because there's more dialogue, there's more space, more time for dialogue, there's better audits of data, better tracking of data, lots of good examples of that here. So yeah. With without taking up too much more time, I wanna, hand it over to to Graham who's going to, share a bit about his work with Claude Code. Thanks so much, Daisy. Yeah. It's just incredible to see the the number of questions flying in from everybody across the world. So, So, again, just tremendously honored to be here. I'm your opener. So I'm gonna get you guys really excited. You know, I do a lot of screen casts about this stuff, AI, building, talking about generative AI and technology on LinkedIn. And and I do this because I really want to, hopefully, inspire and turn on your imagination. You're gonna see incredible, complex, really, products and tools from my colleague, doctor Nashipko later. But I think my goal is a little bit simpler. I wanna make this less intimidating. I think there's a lot of people already building, but I wanna, lower the activation energy that it takes so that you think, hey. I actually could build something. Maybe you don't have any experience with coding whatsoever, and that's what makes Claude Code and these tools so incredible. You don't need to anymore. You know, I've built tools for twenty years, through MD Calc, and I really think the people who best understand what medicine needs are the people doing the medicine, practicing the medicine, seeing patients day in and day out. Hopefully, that's that's probably many of you, on the webinar today. So, you know, I spent years learning to code and troubleshooting bugs and dealing with all the other problems that came with programming, but I actually don't think that's the problem anymore. It's not like, can you build it? Can you learn how to troubleshoot bugs or what technology to use? It is wondering what is worth building to you and for you and for your patients. And you have an advantage that I never had. You have clogged code. But the other thing I wanna just, you know, level set as well is not every problem in medicine has to be a big Swiss army do everything tool. Sometimes you just need a pocket knife. Sometimes look. I know in medicine, we often have just, like, little pain points, little annoying things that we have to do, and, like, a simple tool can, alleviate that pain point and get you moving so you can do the next thing you need to do for your patient. You don't necessarily need to try to replace the EMR tomorrow. But if there's something that the EMR does poorly or doesn't do the way you want it to do as the expert, as the professional in the room, you can build something that Claude Code can do. Right? It kind of changes what the rate limiting steps are. It's no longer like, can you troubleshoot bugs and learn how to program HTML or JavaScript. It's do you understand the problem well enough? Well, I think most experts do. Does the tool actually solve the problem? Can it be trusted? I think trust is gonna be a critical piece in the future. And then if it works, can you scale it? Which is arguably a different question entirely. So these are hard problems, but they're problems that I know you're gonna solve. If you're a nurse, a doctor, a social worker, a pharmacist, listening and watching, you know what your problems are, and I think you are the best one to fix them. So, after my little soapbox and diatribe there, I will, I'll jump in here. I'm going to show you a couple of demos. Again, my goal is to get you excited. These are not going to blow your socks off. Mihal's will, but I don't think mine will. I'll just tell you why I built them and how I built them, and I'll show you a little bit about Claude Code if you don't know about it. So I'm just gonna share my screen here. That's a real big picture of me. Okay. The first thing I will just tell you, and probably many of you on the webinar already know this, but there's two versions, two ways to access Claude Code. One is on their desktop app. I'm on a Mac here, but you can see that, and I'll show you this demo in just a second. But you can see I can go back through and interact with Claude Code in a visual way, and there's a preview here on the side that we'll go to and take a look at. For people that really have no programming experience, this might be the easier place to start because it gives you kind of a a nice clean, user interface that maybe you're familiar with Claude as a, as a generative AI tool that you can chat with. And this is a similar interface. And then I'll show you, what I use actually way more often, which is, just like legit, legit OG Claude Code, which is just, in the terminal here. Now this might look scary to you. To me, I grew up I'm a I'm a Xennial. So this looks really familiar to me, being a, a terminal interface. And the way you start it is just type in Claude, and you go to the directory you want. And now there's a little Claude icon, and now I'm I'm in the tool. And there there's lots of information and tutorials on how to actually install this at this level. There's tons of YouTube videos on how to do this really easily. But let's go back to the demo here, and I'm just gonna show you, what I've created and why I created it in the first place. So, we'll start here with, this, well, baby nurse note generator. If you're a pediatrician, a neonatologist, you know, the tools that that many times we use in clinical medicine are, one size fits all tools for everybody in medicine. Right? I'm an emergency physician. Hal's a cardiologist. We're all given, like, one note generator to use for all of the different things we do all day long. Mahal is doing caths and putting stents in coronary arteries. I am doing resuscitating critically ill patients or evaluating patients for trauma. A pediatrician, managing you know, taking care and seeing well babies in the nursery has fundamentally different work to do than what I do as an emergency physician or a pathologist or a social worker or a pharmacist. And so the idea here is you can take the problem that you have, which is like, hey. I don't like the way my I have to write my notes, and you can build a tool that makes it much easier to do this. I built this in, you know, three minutes with Claude Code a couple weeks ago. You can see it's generating a note here on the right side. It's got some of the features that maybe a a pediatrician needs. Like, oh, here are the APGAR scores that are gonna get populated on the right. Here are the vital signs. There's some screening. But I also wanna show you how easy it is to modify this tool, and make it better. I was talking with my, my friend, Michael, who I inspired to try out Claude Code yesterday, and he said, really, the way that he is starting to use these tools is build a really simple tool and then start to add on features to it, as opposed to building, like, the the perfect vision for the tool, which you can get better at over time as you start to understand, how Claude code works. So let me, just pivot back over to my Claude code window. Because, really, the idea here that I want people to take away is building with, Claude code. The best way to do this is just to get started. Do not get overly intimidated. It's okay that you don't know how these tools work. Just give them a try. So let's say, hey. I want to add additional screenings for the screening, for diseases that they do in, I don't know, Saudi Arabia. So Claude is gonna analyze and look through that WellBaby nurse generator. I'm gonna give it permission to, look through all of my code really quickly, and then we'll see. It's gonna figure out, what they screen for in in Saudi Arabia. It's different than what we screen for, here in California. I do know that because I looked it up ahead of time, and we'll see what it comes up with. So as this is running, I'm gonna pivot to my second demo, and then we'll, we'll go back. So, one of the problems that, if you are in emergency medicine or hospital medicine or you just see patients anywhere, you understand how complex patients can be over time. And as we're generating more and more data, it is really challenging to keep up with all of the clinical notes and all of the data coming at us as clinicians. If I go through, this example here, I've created I had Claude, go through and create a bunch of fake sample, notes, from this patient, who delivered a baby but has a postpartum fever. Here's some nursing notes. Again, Claude generated all of this for me. Here's the the Sarah Connor's history and physical exam. Here's more nursing notes. All of this information. And the the problem I was trying to solve is, hey. If I'm the ER doctor seeing this patient who is kind of what we call a bounce back in The United States, who's come back to the ER because she has a fever, after she delivered a baby, it is really hard to assess all of that information. And so I asked Claude. I asked, hey. Maybe Claude could help that. So Claude built two things for me. Hey. Maybe the patient wants to understand what's going on with her. What what what's going on? Why does she have a fever? And so from those 25 clinical notes, I had Claude build a patient summary about, what's going on with her health so the patient can understand why she keeps getting these fevers, why she's coming back and forth to the hospital. But I also had it build you know, imagine if I see the patient and I'm not sure which antibiotic to give her. I also build, like, an ID consult. Imagine you're an infectious disease doctor, and Claude could build out a summary for you when you're about to go see a new consult. Say I order a consult, and Claude now builds an infectious disease consult. So, again, this is taking the exact same data from those clinical notes, and it's just reprocessing it and reunderstanding it from different two different point of views. One is the patient, and this would be an ID doc who cares about certain things and doesn't care about other things as much. Right? So cares about, some abnormal vital signs and elevated white blood cell count, cares about that, oh, this patient had a positive urine culture that was resistant to a bunch of bugs. Like, the ID doc is gonna care about that. So it's highlighting and surfacing all of that information, so that the ID doctor can make the best recommendation for the patient. Again, my goal here is for the ID doctor not to, have Claude do this work, but to surface the information, so that it's, easiest to find. I'm just telling Claude to keep building here. And I'll go on to my, my second demo as I switch my sharing back. I'm seeing a ton of great questions in the chat. This is what I like to call dude, where's my inhaler? My idea here, again, your imagination should be the limit with these tools. My imagination, came to me one day and said, wow. Wouldn't it be cool if my patients with asthma and COPD and maybe their parents, right, my pediatric patients, could know if today is a good or a bad day to go outside and play? I'm in Northern California. Obviously, we we sometimes have wildfires, and really bad air quality here. And so I imagine this from three perspectives. Again, understanding different needs using the same data. And I said, hey. Could, could I just is there any publicly available data sources that will pull in air quality data? And then how could I use this for a patient, a physician, and then what if I'm a public health authority? So I'll just show you. I tried to also do this generationally. So I've got a Gen z patient. I asked Claude to turn this into what Gen z might wanna do. We'll we'll take, my my hometown ZIP code of Overland Park, Kansas here. Shout out Overland Park. And, oh, look. It's saying Overland Park. This is real data, live data, that that Claude is pulling from another API source. Oh, Overland Park's got great data. Oh, I had to build a testing mode, so I could actually see what changes over time. Oh, as I make the air quality worse, it actually gives different recommendations for people on what to do. I can change over to a millennial, who's managing maybe a a cohort of patients. Maybe this is a pulmonologist who wants to load in all of their high risk patients. So, again, I had Claude build high risk patients for me. And, now we're looking at, say, how how's Manhattan doing? Manhattan's got some air good air quality right now, but it's gonna change over time. And you can imagine you could sort these by severity. You could look at low risk or moderate or high risk patients, and then you could even click a button to call them. Right? And finally, on the public health perspective, we'll load in some data from New York State here. Maybe I'm a state regulator. Maybe I'm the New York State Public Health Association, and I want to see where the hotspots for air quality problems. Well, this found every hospital in New York State. We can zoom in here to New York City, and this is live air quality data with predicted poor air quality over the next, I think, twenty four hours. So you could have a targeted approach of where to, where to potentially help patients. I know I'm, at the last few minutes of my time here. I'll go back to, Claude code, which is finished now. And I'll just show you that really quickly. But the idea here is everything can be optimized and customized for your unique scenario. You know, we have not had that option really ever in medicine, that based on who you are or what you do or where you are or what language you speak, you can change things and customize things to get them the way you want them, the way you solve problems. And you can see here, we asked it for Saudi Arabia. Saudi Arabia, screens for a lot of conditions here, and so it's added all of these additional screening tests for newborns. You know, I I know that, Saudi Arabia tends to have a a larger population with g six PD deficiency, so I I know they screen screen for that ahead of time. And so you could imagine this tool could be customized for anybody. Maybe it's, maybe it's a pediatrician that's not in the nursery but is seeing patients at a one month visit. All of these tools fundamentally let the expert, which is you, let the expert be in control of what technology to use, how to use it, and turn things into exactly what you wanna see from your technology tools. That is one of the biggest takeaways. We have not had this opportunity ever in the history of medicine since the development of the EHRs. We have not had the ability to build customized tools for each of us or customized tools by specialty. And, finally, we have the ability to match what the expert sees and how we understand the world and the problems that we need solved for our patients, and then connecting that with the ability to actually solve those problems with a user interface and code that works. So, hopefully, you can tell I'm super excited by this. And, in my last, my my last minute here, I'll just, share one final reason I like, these tools so much. Again, this code is just being built on my computer. If you use the Claude desktop version, it actually can build it, and push it out to the the web for you as well. It is an iterative process. You get better over time. So the the biggest, hardest step is just installing it and getting started. After that, you get to play and use your imagination. Claude has guardrails. It even has commands and tools that will help you start to do things that are obviously really important, health care, like security checks, and it understands what you mean. You can give it, not what is a perfect prompt, and it often will infer maybe what you want, especially if you give it a little bit of context like, hey. I'm a physician. It's gonna build something different than I'm an electrician or I'm a student or something like that. I will pass it over to my new friend and colleague, Michal, and I will pop off stage. I'm really excited for you guys to see what he's built as well. Thank you so much, Graham. That was incredibly exciting. And, well, deep inside, I'm actually an Andy Karl groupie. I, I followed the project since the very beginning, and, I'm I'm really happy to see where it's going right now. It has evolved tremendously. Let me tell you a few stories from my life, which is a little bit, you know, in between medical and informatics. So as I told you, since about twenty years I called, I think I was just before med school when I started doing it. And, all of my frustrations in medicine, I tried to translate into working solutions. There were many of them. There were electronic health records. There were drug search engines. But, something profoundly changed when AI, allowed us to help with code. So about three years ago, when we saw the first iteration, so we have code with Copilot, then there was, there was Cursor. There was Windsurf. I started using it. But last year, when Claude Code into being it came into being, actually, in February, it was the first anniversary of Claude Code. I attended the the the birthday party. It was incredible meeting Boris and all the team. It accelerated tremendously. And whilst the terminal can be a little bit intimidating, I think the team has done a really good job to actually circumvent it. And I'm gonna show you a little bit more of this. So what I'm going to do, is I'm gonna show you three demos. I'm gonna show you previsits. I'm gonna show you, NetDuty's system for solving problems with organizing on call duties. And, I will also, show you, a system for, which which was created for the hackathon, so the post visit. So without further ado, I will just stop sharing and will share my window with with the system. So first of, of the solutions is called Previsit AI. So, basically, this is a system that I have created about three years ago, when I was head of department, in in Belgium, and I was a little bit frustrated by the process of, repetitive, anamnesis. So when the patients came in, they, rarely remembered what medications they were taking. You know, my wife knows red pill for the heart and so on. So I tried to experiment with LLMs a little bit, and this is a huge system. I'm gonna just show the standard demo of the cardiologist. Before to visit, the patient receives a text message, and then he clicks on it, and he discusses with a very polite and very empathetic, AI. So this is the assistant, like, hello. Can I ask you a few questions? The the the consentment is absolutely crucial. So is it your first visit to doctor Anthropic? Well, I've seen him before. And here is a role of a cardiologist, which can be customized. Can you tell me when was your last visit, like, a year ago? You can make all the mistakes you want. Are you experiencing any problems? So this is a cardiology visit, so I go with chest pain. So do you have any current concomitant diseases? So patients usually don't understand those tricky terms. So concomitant, you can ask, oh, for concomitant, I mean, you know, other medical conditions like diabetes, hypertension. So I have high blood pressure. Okay. Since when have you been diagnosed? Like, three years. Thank you. Now could you tell me about any medications that you are using? So it's usually standard answer. My wife knows that. But, could you please ask your wife? Oh, she says that, telmisartan, and amlodipine five, and so on and so on. And the patient sees the progress bar. He can actually converse using voice. There's also, like, modules to import the medications. And what happens after this is the doctor will receive a medical grade summary. So, basically, this is the translation between the patient's language and the doctor's language. The goal is not to replace the anodesis. The goal is to enhance it so that when the patient enters the room, we see more or less what are we dealing with, and, also, we can validate. And all those informations that the patient does not have, in the room, are already there when he interacts with AI. It it is really a huge system. So thanks to Claude, I was able to basically explode it, create, an interface for API for integration with other systems, documentation. It, it it it's it's absolutely incredible how much work it it allowed me to save automated translation for multiple languages. It was really, really great. So the second system, is called med duties. And this is a thing that I've created, like, eight years ago. You know? It was a very old technical stack. It was, like, old PHP. And I it was really well, very utilized, and I think it was, like, in about 50 hospitals right now. We have about 2,000 users, but I had a problem with scaling it. So what it works like so, basically, it solves the common issue, like, you know, you you need to be on call. And the the way that it is usually organized is that there is a secretary or somebody who's responsible who gets the the the staff's wishes, and they can go through post it notes or they can go through, through Excel, and, basically, whatever you're going to to fill in, it it won't go your way. So here I have created the system that is able, like, to show, okay, I want I don't want this weekend, but I can sacrifice for the department, and I reluctantly can take, like, the eighth or the ninth, and so on. And then the moment that you are going into planning, then you see that doctor, doctor I think that for July, we didn't put any wishes, but on June, we have, for example so doctor Walker cannot do fifteenth, sixteenth, or nineteenth and so on. But, if, if all is going to work good, then there's this special button to generate according to the wishes. And all of it connects also to external calendars, to Google Calendar, to iCall, you have all the stats. You have, like, equilibrium. You're able to do so much more. But, you know, I had an issue with scaling it, and that was, that was a problem because it became huge. It was a large code base. So I put on Claude Code to, to help me out a little bit. And Graham showed you the terminal, and I was using the terminal until now. And about a week ago, I came into what the Claude Code team did with the with the redesign. So, basically, code is integrated in your standard, Claude application. But before, it was, I would say, quite basic. But what is happening right now is that you have multiple modules that are integrated with it. So, like, you can see the preview of your applications. And it's not only preview. Claude actually sets it up for you. And it before when you were coding, even with Claude Code, the GUI, you needed to go out of the application, like, inspect something. And, like, here, for example, I see that this button here, it's it's a little bit, you know, misaligned. I will I will check it out later, and I will show you what it is. But long story short, what I wanted to do is I wanted to give some conversational layer to my to my, application. And within twenty minutes, I have asked him to generate, to generate an assistant that went into the whole data structure. It understood my application, and I was like, give Graham all the Sundays. And then, he identifies that, oh, there is no postal code. Ah, I'll give Walker. I'll give doctor Walker. I think I didn't give Graham the name. Okay. So which columns? All of them. Alright. And now he understood, okay. I understand that doctor Walker no three columns. No. All of them. And then okay. Assign only to night shift. Sometimes it works. Sometimes it doesn't. Okay. And then here, it assigns doctor Wako to all the night shifts and the rest we can fulfill with the rest. But long story short, in about twenty minutes, I have created a completely new interface, a conversational one where we can actually talk to medical duties, and we can say, like, you know, I don't want this Sunday. And it's it's actually how those duties are are are are usually filled in, like, with, with all the system that we have currently. So what is happening here is also quite remarkable because you don't have to go out. So whatever you're going to do, I will try to see maybe if it's going to help me fix that. Yeah. So here so this thing overlaps with the bottom filter bar. Can we move it up? And so on. A few tips that I will give you while it's working. As Boris said, always use the best model. Okay? So if you have a possibility, always use the best model and use the most effort because this way, you won't be you won't have to reiterate. There's gonna be less room for error, and it's gonna just end up with the net, benefit for you. So there are many other, elements in here. Like, you can see your files. It prepares the tasks. It prepares the plan. It create it shows you all of the diffs and what you have created, all the, all the files. It's it's it's actually not really an IDE. I would I would call it a completely, you know, dynamic subset of self generating software. It's really incredible how well integrated it is. So I I think that I'm starting to use GUI much more than than I use terminal recently. And this, this is very important. So if you wanna accelerate its achieved code, you usually use a parameter in terminal which is called dangerously skip permissions. Remember that this works super great, but there might be some hiccups. Like, for example, when I was doing demo for post visit, I asked it to create, a special server in The US. And I had, like, this API for my for my server provider, and he said, woah. You already have five servers. I think you should delete the one that is inactive. Can I delete this one? And I said, like, yes. Okay. And then he's like, oops. Sorry. I deleted the wrong one. So always make sure that, you know, like, what are the limitations. Hopefully, the the there wasn't an issue. Oh, see? It already helped me out with, with the button. So this is done. Alright. So I will show you the last demo. And let's go back here. Alright. And this is, this is my hackathon demo. So post visit is a system that revolves around the problem that happens after the patient leaves the room. Basically, they don't remember what medication you prescribed them, what is Mediterranean diet. Long story short, there is a big hiatus between what is going on, in the visit at the visit and the next visit. So I wanted to give my patients an idea, a tool that will help them. So it revolves around, the entity of the visit. So first of all, you are able to see your last visit. You are able to ask about all the, important aspects. It explains to you all the terms. Each of those elements, like, for example, doctor's recommendations, will allow you to chat with it. And, all of it revolves around a scribe. So, basically, as you know, currently, one of the most important advancements in in in making doctors actually more more happy during the visits is the AI scribes. So I wanted to give an AI scribe to the patient that will allow, them to, record information. And, also, on top of it, he has possibility to connect his own vitals, to connect it to his lab results, to to upload any kind of medical reference, and also, guidelines and evidence. So very important thing is that he can also connect to existing sources of truth. There is also a doctor panel in here where if the patient wants to contact the doctor, he is able to also investigate whether the patient needs to be seen or not. So, this is a hackathon demo. But, coming back to the process of creation, what I wanted to share with you as a tip, when you create a big project like this, the most important thing actually is starting up with something which we call a PRD, so product requirement document. So when I was told we can code right now on the hackathon, I started off by creating a very large document. I, of course, use Claude to do that, and I described very thoroughly what I wanted using voice. I used Super Whisper. And then after two or three hours, I was like, okay. This is huge. This is gonna take days. So I went to take a shower. I went to brush my teeth, and I came back to check on it. It's like forty minutes, MVP was absolutely done. So if you create a PRD and you detail it very, very well, then this is, just, make wonders. So, we're gonna talk more about Technicities during, the q and a, but I wanted to end up with some, remarks on why actually I built using Claude. So for me, it's it's light speed what what is happening. It it it feels basically like pairing with someone who really understands the problem. Right now, there's a Hopeless four point seven hackathon. I can't wait to see what those guys are shipping. And when you work in the directory, you basically have full access to your project. So Claude will also enable you to set up not only, your code, but also your servers, your, external services. And, basically, it's an adventure to see how fast the team is shipping. Like, literally, during the hackathon, they ship three essential functionalities that I actually needed, like, parallel agents, and it was crazy. So, as my last slide, I will share the toolkit that I use. So I think that I have to change instead of terminal that I will be using more of the GUI right now. But if you use terminal, don't forget there is WARP, which is, like, AI enabled terminal. Very good. And I think that team loves Ghost TPI. And browser, of course, it's Chrome because it's connected. You'll also need an ID, to to review your code sometimes, but less. I use cursor for some time. I love Raycast, which is like a Swiss army knife where with the common space, you have access to all the tools. Dictation is so fast, much faster than typing. I actually use a cheat code because I'm a native Polish speaker, which has the most token density and context security, so I cheat in there a little bit. And, this external DPS, if you wanna, like, offload your work inside the Claude, then you might use an external DPS server. But Claude right now actually has a possibility to configure it from within the application, which is really great because you can just go for a walk and, you know, go use it from your, from your phone. So that is it from me, and I think we're gonna, take it back to Daisy. Thanks, Mihal. That was that was really great to see. I'm gonna I have, like, one slide on the Claude platform, here, and then we're gonna jump into the q and a. There's so many good questions here. And and, like, Hi, legitimately, Gun. also so many good ideas and suggestions and and discussion. And and and really, Yeah. Definitely. I just wanna say my heart is so full seeing so many. people who care about the world. I actually like, so I have a slide up on the screen with, our stack of products and services that you can use, and, like, Claude Code is on there, which is obviously the one I. work on day to day. But I I I really wanted to take a moment to speak directly to the practicing. physicians way, in the audience. we're synced we. have a broad audience. I'm the child of a pediatric neurologist and an interventional cardiologist. I grew up around medicine. My parents both did really important research early in their careers. My mom did a lot of work on, epilepsy in infants, and my dad was part of the team that discovered angioplasty. And they did they both cared so much for the world that I think they got tired of research and wanted to interact directly with patients. And I suspect that that resonates with a lot of people in the audience, and I really wanted to take a moment to speak. to that because I I think that a lot of the things I heard around the dinner table growing up are really barriers that are being broken down by the ability to build anything. Like, we saw some we saw some really good presentations, about, like, how to integrate AI and and and integrate generative, chat into, products, but, like, it doesn't have to be that. Right? The it's it's you can build products for your patients that use AI, but more importantly, you can build products for your patients, period. And they don't have to use that AI, but, like, you have the ability to build what what you feel like your patients need. And I think that this is a really unique and important moment. And this is this is, like, what gets me out of bed in the morning, really, seeing the difference that this can make in the world. And and it makes me really excited and proud to work on Claude Code. So, and and a lot of the questions reflect that. And so I'm I'm just so excited to chat with you all and and follow-up on a lot of these things. So I I wanna bring the the both the speakers back on the stage so we can start answering questions and have some dialogue. Hi. Do we wanna take the slides down? I think people can see our faces. They'll need to see our faces on the slide also. Yes. Awesome. Okay. Am I supposed to be I'm supposed to be moderating the questions. I think I I think we have people filter. It's hard because it's incredible seeing these questions, you guys. I mean, I I'm not sure how much, gosh. Yeah. the audience can see, but there are hundreds, like like, you know, 50 every one minute. And so it's, there's so many great great questions, and I'm just really excited. Yeah. I I'm really excited just to have so many to have Anthropic see so much excitement and activity from health care professionals, nurses, social workers, pharmacists, physicians, NPs, PAs, because that's gonna get Anthropic excited to build more stuff for us in health care, which is what I want to happen. Yeah. So, I mean, a lot of the, questions started with, like, how do you make sure all of this is HIPAA compliant? The European version of that starts with a g, and I'm not as familiar with it. GDPR. Yeah. Yeah. Fun different story. I was HIPAA trained for the first time, I think, when I was 17 or 18, so I I definitely grew up around medicine. How but I'll ask you two first. How are how are you handling HIPAA compliance? I know you you have, like, pre generative AI experience in. software development. How do you see that changing? How much technical expertise do you actually need to make sure these things happen versus, like, Yeah. I will pitch, in on the European that is kind. of logistical? So developing. in Europe is is really difficult can take, it comes to regulation. HIPAA, You know, I live in the heart of. regulation in Brussels. know, Daisy, So in Europe, will say have, GDPR, which is about, first of all, data. like, I really it's basically a set of policies health you have to introduce so incredible. on the on really appreciate. that people set care about privacy. of use, Like, people want to follow, both regulations, course. We have all don't think they wanna, just follow, on top of those. bottle think they want, you follow, them because, they respect, that, that the world, is making fun of us. deserve privacy, generally, about their data and their health, forget, that in co, work and clothes, you have a legal plug in, care workers. and you can I love everyone so much. about how it should be implemented. know, I I think it's also couple things. One, signing with providers. I would imagine everybody the legal this. call, appreciates about encryption and basically building with privacy first is. very important. But in Europe, is before the Internet really called the AI. Act. That is why AI Act is a quite stringent in, regulation every, where the solutions, there are thousands of fax machines health care, are being put on top of hospital. the so called risk pyramid. I would argue that, unfortunately, it requires, HIPAA could lot use a refresh. in terms of, when I was, creating my startup, protections basically, along with, better reasonable, percent of the budget should go to the lawyers to basically best of both worlds. So I this. I will say that. first. There's will also just say that, and I AI and won't I. won't speak for, So know, it's quite code. or anthropic, because I'm I don't speak for them, but I I'll say that, we're slowly know, HIPAA is a set. of standards nice expectations. It is not necessarily scene developing not the same, too as, like, but I think that, compliance. in? US, HIPAA a little a set of, straightforward. So for example, standards visit, that, are was mixture of, like, with HIPAA Pathway from the beginning, and security is also auditing signing, and record keeping. legal. And so we ask say, HIPAA then don't forget that if you wanna implement it, means different things engineers different people. And so, to do, like, you production, fundamentally, ready code within for this version of HIPAA from 1996, it's one thing creating something need your computer, to talk with a a HIPAA compliance actually, running it with live data of patients, anything especially health you're within an institution. that does touch this actually data for be bunch of reasons. run through your would. argue that, you can get really far because you're gonna build the tool that your colleagues and you want to use, and you can take that, that adoption and that usage and then help build, help build a case and an argument to say, hey, health system or employer or pharma or whoever. Hey. This thing is so good. Help me figure out how to make it HIPAA compliant. Help me figure out how to integrate it with my my system of record, whatever whatever that is. Yeah. I can tackle this. So in Europe, basically, Yes. it's worse because you're basically dependent on goodwill of the provider of the EMR if they're going to provide an API. And my life struggle was actually, you know, growing a project and then just jumping, I think with doing head first into a wall. good points. But in, US, if you want HIPAA, compliance, which is called USCDI, pay someone to make sure that, it happens. But I think we can get you provide you with lot of the way there if you are. developing a solution. additional tools. So, We are, working is, on, quite possible to integrate it. part of the the follow-up to this event, not working on releasing, a HIPAA plug, in. but in US, I'm, guessing have, this means, we also need, to do a GDPR plug in where where share the will go through within the HR. audit everything that it can. I certainly would not take that as legal advice, but it maybe would speed up the HIPAA audit when you actually. have a professional do it. But this this really gives you the ability to get yourself 90% of the way there. I don't think we're really quite yet recommending that physicians, like, directly ship the applications they build to their patients in the practice. without asking does great job at two people. APIs Maybe we'll be there in a year, it. but, I think it's worth just, you know, there's a lot. of nontechnical barriers in the way of that, I would say. I I think you're you're totally right, Daisy. The the idea here is, not that, at least today, not that Claude is replacing HIPAA compliance, but a a HIPAA compliance audit review with an expert that might, take two months and whatever, $10,000. If you've already built it privacy first, like, Mahal has said, and then you run it through, you know, the the HIPAA evaluation skill, then maybe that brings it down to it's a two week engagement, and it's a $2,000 engagement. And so you just speed up the process, and so you're able to just do things way faster and way cheaper than you could. I I think that's the right approach for now. Yeah. Let's, jump to let's see. There's a lot of questions about how do we connect things. Right? How do you connect your note generator to EHR? I assume similar questions for Epic or, what's the third one that people might be asking about? But in general, I will take maybe, what is this, to kind of incite a revolt. how does all, of this the EHR is a problem, ecosystems, maybe just create your own. does it? Like, there's always a legal problem around it and, you know, there's payment issues and whatnot. But if the the doctors will actually go first and they're gonna show the industry what we need and create their own solutions, then the change will come. So I had a lot of talks with, you know, doctors from recently a, a radiologist from Austria. And she was like, he created this great solution, and the EHR is not letting me in. Just create your own. Try it. Like, see if it's gonna be possible to implement it, and this is gonna force the industry to change. Is that a thing we should be building, plugins or skills around? Sounds like it would be very helpful. Like, I I do think these are definitely in the realm of things that that Claude is likely to understand as well or better than most developers, in terms of is is it just this very niche information that's very specific but not very broad in scope. Yeah. And I think the industry is moving in this direction, Daisy. There's a there's a, a law policy called TEFCA, which, is is going to be, forcing more, APIs to be open and accessible, you know, with obviously some some security and, appropriate guardrails. But I would also call out just the the the, the listeners, and people watching as well. Like, this is what we should be, demanding, expecting from, anybody that houses the the data as well. And I think that is where people are gonna start asking these questions. Well, why can't I access this data as long as I've done training and it's secure and it's HIPAA compliant? You know, all these things. Why can't I. connect my thing that would help me take better care of my patients? I think patients are gonna start asking this as well as health care, professionals, and that may actually drive this conversation and move this faster and further as well. If multiple stakeholders are saying, well, why can't I access this data? Because it's gonna help me make better decisions for my patients too. It's a really interesting time. Yeah. I couldn't agree more, Michal. I I think as more people are building things, I've said this for a couple years now. Like, health systems and EHRs, and everybody's gonna start to have to ask, like, a really difficult question, which is why aren't we letting our physicians build these tools? Why aren't we allowing these things to connect? Because fundamentally I'll get back to my original point. Like, the experts are the ones that understand what the problems are, what the pain points are for them and their patients. People are just gonna start asking questions. Why isn't this available? Why is this a problem? So I'm excited. Yeah. I mean, I I really hope that this goes the way of laws that protect physicians when they say see an emergency in a public place. Because, like, the reality is having physicians be able to build on these things, be able to see data, be able to process data at scale is really gonna save lives. And I think we need to take a very hard look at what's standing in the way of people saving lives. I have one question that I wanna we've seen, like, 10 variants of that I wanna read directly, which I think is is like a very good discussion. We've actually been talking. about. it a little bit backstage. I know. I just wanted, to say that, person says, technically, in one week, I went from difficult to distribute a terminal you can just building, ask ClaudeQuote, two it will ship on a public domain, my colleagues are starting to use. Yeah. will ship on a public server, challenge this this, will, work. so fills you're treating. with patient data, will, then you're, exposing it at the least, be happy for a. week just thinking about that first sentence. can set up VPNs. My, for example, is distribution. it can set up Tailscale, Right now, is an sending HTML files, via OneDrive. set up a local network where it is used only a position computers. to take a, Claude built kind of, tool of of complexity, it available would definitely practice or department, it to somebody in the institution. some users if you wanna do it locally, at all? if there is no patient data, if think this is you're afraid that's gonna question, be leaked, like it's only. And I think project with anonymized consensus, then just ask Claude how to do that because it that. does it really well. And, you know, I don't know if it's anthropic or not. I think it'd be really cool if it we we did. But, I think that figuring out that last phase, is is super important, but I wanna get your thoughts on on distribution. I'll say for that specific question, Daisy, this is you know, I'm I'm also a clinical informaticist. This is where I think sometimes physicians don't know who to talk to either in their health system or their medical practice group or their organization. So that's gonna be, you know, often your chief medical information chief medical information officer. That's gonna be, typically a physician, a clinician who also wears an informatics hat like I do, and they're gonna be the bridge between clinical operations and technology. If you don't have that type of person, it's probably like a CIO, CTO type person, or it might be a chief medical officer. But I would say, you know, get fifteen, twenty minutes with that person. Show them what you've built. And, if they are a future thinking visionary person, they're gonna see the value there and say, hey. How can we make this happen? Or maybe they're working on something similar in, the pediatrics department that the cardiology department doesn't know about. So connect with people in your organization who both have the vision, connections, network, and the power to approve this stuff. And again, this is gonna become really common. This is gonna be a problem for CMIOs to figure out how to address. It's like, oh, I have 50 doctors building really cool stuff. How do we how do we get it so that they can deploy it inside our system safely, appropriately, and securely? Yeah. I mean, there's so much of this. Oh, go ahead. Yeah. I'll start. Yeah. I I also wanted to conclude by saying thank. you so much. I already told Boris that numerous times. But what you have achieved with with Claude Code is is is nothing short of remarkable. So, you know, before the outlet for creativity of doctors was, like, creating publications and waiting until somebody maybe will materialize that. Right now, we have, like, this this incredible tool, this incredible machine that allows us to prototype in the matters of hours, sometimes minutes, not months. So, I I'm incredibly excited to see what's going to happen with this new wave of of physician builders. I completely agree. There's so much of the energy from this talk that is just the everything I've wanted for the world. Thank you very much. worked in programming you. languages. for fifteen years, and part of what I wanted from that was just simplifying the workflow for non programmers. And and and, like, that whole dream has come true in, like, a year. And so I'm I'm thrilled to see people showing up here talking about the prototypes they've built, building things that really are gonna help their patients. I just wanna thank you both so much for taking the time to talk. I know we didn't get to a lot of the questions. There's there's literally hundreds of them. So I I I hope we will get some time to to follow-up, in in text form or something like that. I will, do my best to push that. And, yeah. I suppose I'm supposed to to wrap up now. I'm getting a red banner that says overrunning by eight seconds. I also know how tight the timelines are for doctors, and I I really genuinely appreciate all of you who took the time to to listen to what we had to say and to be a part of this change. Do either of you have any kind of final words you wanted to leave people with? Thanks, Stacey. Yeah. Yeah. I'll just say, you know, thanks for, thank you for building Claude code for making this this possible. I'm really excited about the future. I think the future is, certainly more unknown in a lot of ways, in today's world right now, but I like to think really positively about all the, great problems we can solve and all the great ways we can help take better care of patients. Thank you, everyone. I am not sure that I have the controls to end the the session, but I think the people backstage will, and, look forward to seeing all of the things you all are building, the difference you're making in the world. Thanks, everybody. Bye.