How to Build a Defensible AI Startup – With Dr. Thomas Kelly from Heidi Health

0:02 Georgie Healy: Founders scale faster on Deel. Set up payroll for any country in minutes. Hire anyone anywhere. Get visas handled fast and get back to building. Visit deel.com/dayone. That's d-e-e-l.com/dayone. What are some critical decisions you made when building Heidi from a technical standpoint?

0:29 Thomas Kelly: If I have to, as a doctor, like, I can't trust the output and I have have to review like every single word of every single note and like key facts are wrong, then it very quickly kind of stops being useful.

0:39 Georgie Healy: The year of the AI agent. Agents, agents, agents. They're changing our lives, Tom, or are they? What do you think they can do and what can the everyday person actually get from an agent?

0:50 Thomas Kelly: If Heidi's valuable, I as a doctor don't have to edit it very much. It basically should read my mind and write what I would've written in the room.

0:58 Georgie Healy: What kind of AI startups may not survive due to the type of problem they're solving or a very outdated approach.

1:05 Thomas Kelly: All B2B SaaS. Yeah.

1:09 Georgie Healy: What's that must-do if you're building an AI-powered company like Heidie Health? Hello and welcome to In the Blink of AI, your weekly front-row seat to the AI revolution. I'm Georgie Healey, and this week I am speaking to Dr. Thomas Kelly, the founder of Heidie Health. They're the AI startup that's growing faster than Canva. And on this show, we talk about why accurate medical notes are crucial, the future of how health records are interpreted, attention hacking and personal brand as an AI founder. We also play an AI game. And if you wait till the end, you'll get Tom's surprisingly hot takes. I know this is going to be a fan favorite episode. And we're keeping up momentum on the show because next week is our first ever in-studio recording with a multi- millionaire founder. The questions I asked actually made me uncomfortable. So subscribe now and it'll be waiting for you next Friday. But for now, I can't wait for you to hear from Tong from Heidi Health. Let's dive in.

2:17 Thomas Kelly: You're listening to a Day One FM show.

2:20 Georgie Healy: As a startup founder, you're juggling multiple priorities from the expected, like finding product market fit, to the unexpected. Like customer requests for SOC 2 or ISO 27001 certification. Achieving compliance is time-consuming, and time spent on that is time away from the needs of the business. And that's where Vanta comes in. Vanta is the all-in-one solution for startups to become compliant quickly and build a security foundation with ease. With a combination of automation, an extensive partner network, and a security marketplace containing 385 plus pre-built integrations, Vanta provides the necessary tools and expertise for startups to achieve compliance seamlessly, no matter how urgent your needs are, and at every phase of growth. Over 10,000 leading companies, including Cypherstash, Handle, and Indebted, trust Vanta to automate compliance so they can focus on growing the needs of their business. Here's the important part. Startup listeners of the show get $1,000 off if they go to dayone.fm/blink. Hey, Tom, thank you for joining in the Blink of AI. I have had you on my bucket list for some time. Thrilled to have you. Look, Heidi Health is a household name. You're on the front cover of the AFR. But just in case listeners aren't aware of what Heidi Health is, can you please give us kind of the elevator pitch.

3:59 Thomas Kelly: Yeah, thanks for having me, Georgie. And, um, yeah, I think I'm hoping most Aussies would have run into Heidi at this point, maybe if they're GP or the physio or something like that. But yeah, Heidi's a piece of AI technology that listens into visits, turns clinical conversations into really good clinical notes. So usually as clinicians, it's really tricky to manage everything in the room. You've got a patient, or maybe patients, you've got trying to figure out what the right question is or what diagnosis they might have, or, you know, basically what to do next. And then you've also got to document everything and create all the paperwork that the patient needs kind of all at once. So very often clinicians don't do all, all those things at once. They end up not writing their notes as they go and then either do them at the end of the day, or maybe while you're in the waiting room reading, you know, 2001 Women's Day, they're, uh, they're busily typing the notes while you're wondering why they haven't called you in yet. . So yeah, I think basically if Haidi's doing its job, we, um, let them click stop at the end of the visit, instantly create great notes. They review them, draft them, put them in their record, and, and that's it. So yeah, uh, fortunately for us, it's been really popular. Clinicians save a lot of time and, um, now being used all around the world. So yeah.

5:15 Georgie Healy: You guys have been incredibly successful when you were a doctor. Talk us through how much of the work was like in that chat that you were having with the patient versus like follow-up and the stuff you talk about, about, you know, um, the write-up of notes and things like that?

5:32 Thomas Kelly: Yeah, I'd say probably about half the clinical time is spent either before or after visits. It's sort of like pretty proportional to the amount of time you spend with the patient. So my experiences were mostly in outpatient clinics because I was a surgical registrar. So we had a lot, you know, the kind of clinics where you go to hospital and wait to see the surgical team. And very often there would be, I don't know, you know, 3, 4 scans that have happened before previous visits that might've occurred, maybe old clinical notes while they're in hospital. So there's a lot of looking through the past and trying to figure out what's going on. And then in the room, every conversation that you have, you at a minimum have to write a clinical note and a letter back to the GP. If there's anything else special, then you might need to create, you know, TAC, WorkCover forms or certificates of capacity, or. you know, you name it, there's different kinds of forms. So yeah, generally if it's a half an hour visit, I probably have 15 to 20 minutes of prep and documentation to do. And very often, because it's so busy and the practice— the clinic is only maybe like 4 hours, you just end up seeing as many patients as you can, and then you do all your paperwork after the fact.

6:43 Georgie Healy: Wow.

6:43 Thomas Kelly: Which is bad quality of care because then you've, you know, you're not likely to remember everything that you said in the first visit. So you're giving, you kind of, you're doing your best, but often you're missing key details or facts of the case. So what we actually find is that Heidi's notes are generally more accurate than the manually written notes because it's not because doctors don't do a good job, it's because they are doing them after the fact. So they don't quite remember everything that happened.

7:09 Georgie Healy: Oh my gosh. I'll listen to an episode I recorded the week prior and be like, I don't remember any of this.

7:16 Thomas Kelly: Yeah, exactly. Yeah, it's like, you know, memory is interesting. Like, you kind of remember the highlights and the lowlights, um, or like strong emotions or like things that imprint in your, in your mind. But I think you remember the key facts of the case and like the main things, but you might miss a lot of the details. And it's interesting, like, I think those are the things that make people great doctors or perceived great doctors. Like, if you remember that they have 3 kids and their names or their ages, like all these little details that you kind of forget. So there's increase in quality, but also for patients, this like perception of just like this amazing doctor that just seems to remember everything about them, which is the best part of it for me.

7:53 Georgie Healy: Yeah, we hear about this, you know, bedside manner, or, you know, that kind of feeling that you get from the doctor. But if the notes suck, I guess it's kind of like, well, I don't really want to just have a chat there. I do want to be diagnosed correctly. And all that.

8:10 Thomas Kelly: Exactly.

8:10 Georgie Healy: Doctors are famous for terrible handwriting. What was your handwriting like, Tom? Did you try really hard to make it perfect?

8:17 Thomas Kelly: I have to say, I was an outlier. I actually had really nice handwriting.

8:22 Georgie Healy: Really?

8:22 Thomas Kelly: Yeah. I think a lot of the doctors that went through my generation, they had, you know, it was like the millennial, like highlighting and like little sticky notes and like, yeah, beautiful notes. So yeah, I was very similar. I had all these colored highlighters and like drew surgical anatomy and pictures and things. So, but it was too slow. That was the problem.

8:43 Georgie Healy: So yeah. Yeah, that would be great if you had 3 patients in a day, but—

8:46 Thomas Kelly: Exactly.

8:48 Georgie Healy: So I was curious about this because when I recently booked in, I got an email notification saying in advance, you know, this session will be transcribed by a— I don't remember if they specifically called it an AI tool, but I thought to myself, I wonder if this is Heidi. Would it be you? And how many Australians inadvertently have been supported by Heidi Health, do you think?

9:18 Thomas Kelly: I think on the consent point, it's really important. We, like, out of the box, Heidi has a way to set up consent. And yeah, we give like patient explainers and forms in the waiting room and all sorts of things. So definitely expect that the clinician is telling you they're using Heidi either beforehand or in the waiting room or in the visit as well. As far as, yeah, I think our reach now, we don't know for sure because we don't We don't capture any of that patient data for ourselves, but we obviously know how many visits we do a week. So we're almost doing 2 million visits a week now. Um, that's around the world. So in Australia, it's probably around 500,000 or so, like a quarter of them. So, you know, over a year, like that's, I don't know, like a lot. Um, you know, 25 million visits maybe. So I think probably like 20, 20, 30% of Australians probably have run into Heidi at this point. Mm-hmm.

10:13 Georgie Healy: That's incredible. That's so incredible. And just on a personal note, you seem to be, you've got quite the follower count. You've got 10,000 LinkedIn followers. As a founder, I'm curious, is that kind of, that's the game now? You do need that personal brand and reach, or are you kind of reluctant and like, I prefer to just stay behind the scenes, or a bit of both?

10:37 Thomas Kelly: Yeah, definitely a bit of both. I think Early on, I found LinkedIn to be a really useful channel for us. So, because often a lot of early adopters or people who are looking at creative careers in medicine or different avenues were sort of on LinkedIn. So they were often— it's a good way to find your first users, we found. And yeah, I think I remember writing lots of, lots of different posts and I'm trying to get people to try Heidi. And then I think as it's grown, generally still very excited to share like team achievements and milestones for Heidi and, you know, interviews like this, because my job as CEO is to try to attract amazing talent and, and like find the absolute best people and make sure they know that this amazing, you know, AI worldwide AI story is coming out of Melbourne and Sydney and you should join. So yeah, that's kind of how I think of it, I guess, mainly for, for employer brand, you know, recruiting, that kind of thing. Not that fussed about my own brand as part of it.

11:34 Georgie Healy: Yeah, the reason I ask is A perspective I increasingly have is that, you know, time to market and, and the availability of these incredible tools that allow founders to, to get up off the ground running faster exist. And so to differentiate, and for, yeah, for brand, having a social presence is more and more important because, yeah, everything around it is democratized. Do you, do subscribe to that philosophy or not really?

12:09 Thomas Kelly: Yeah, I think so. It's similar to, um, you know, a lot of media like podcasts like this, like sort of these, like the loss of centralized media and channels and TV is obviously less viewership than ever. So I think it's similar in that way, like the best, you don't have to do it, uh, it's just an option. So, so there are different channels to get your product in. In front of people. I think for us it's interesting, like doctors are probably typically not a group that you would go directly to. More often people would sell top down or have like some sort of B2B, um, you know, selling to the chief medical officer or the practice owners. Uh, and then the doctors are sort of an afterthought. So I think we're, we're unusual in that way. We, we go directly to the doctors. They don't always use it straight away. Like they might have to get the permission of their practice. But that was a pretty explicit strategy. Having been a doctor, I thought they're humans and they like to use good products and they're the same as everyone else. And yeah, so for us it was important to build a great brand. Like we have good, we even do like a lot of performance advertising, like directly to, to doctors or these kind of like pre-roll YouTube ads of me being dumped with notes and talking about the product.

13:26 Georgie Healy: I need to see this.

13:27 Thomas Kelly: This sounds good. Yeah, so I think, I think for AI products at least, especially AI applications where it's more consumer first, then anywhere where your consumers are is where you should go. So a lot of, a lot of founders go to Twitter or X because there's heaps of engineers of other founders there. So if it's, if it's a software for other founders or productivity tool, that's probably the best place. For doctors, it's a bit tricky, like LinkedIn. Um, there's some communities and Facebook groups and things like that. But yeah, I think it's really only like a seed. You have to have something great that has good word of mouth and then it'll spread on its own. Um, but you can, I don't know, I've seen like some of the companies coming out of YC now and in SF, there's a lot of like, call it like attention hacking, like trying to, trying to have like, I don't know, the most like outlandish, crazy, like scroll stopping. just purely just to get like the name out there. It doesn't even matter what they do.

14:29 Georgie Healy: So have you seen these founders that like raise money from A16Z? Um, like one of them hacked admissions tests for Amazon. You saw this, right? And then they get funding.

14:43 Thomas Kelly: Yeah, that was the one I was thinking of.

14:45 Georgie Healy: Yeah. Right. And it's like, it's not really the way that you would expect to attract, you know, someone trustworthy or something like that. But yeah, this attention hacking is fascinating.

14:58 Thomas Kelly: And I think it depends on your category. So obviously for us, like trust and safety and privacy, like they're key things. So that would never work for us. It would never be something we, we, we did. But, but yeah, it's interesting. Like I think there's no wrong answers. I guess you just want to try to get people using your product first off, just for,, you know, survival at a minimum. Like you need people to use your thing, need to, to pay your team, to, to start growing. And then, and then once, once you reach some steady state word of mouth or groups of users, then, then you can be more precise about what channels or what approach you want to take. For us, it was, and like still today, I still think brand is the most important thing in healthcare. It's very networked. Doctors talk to each other. As you can see, like the number of patients that have seen Heidi is kind of wild. So, so they very quickly— it should spread quickly if you're doing something right. Uh, but what the brand stands for and the way it makes you feel, and, and for us it's like that time to care, like the time to care about the patient, giving patients and doctors more high quality of care, more time back. All these things are just like brand values that, that we try to, uh, I guess do in all our channels, my own, the, the brands everywhere.

16:12 Georgie Healy: Thank you for unpacking that. That is genuinely fascinating and a topic that I'm, I'm noticing increasingly valuable across founders, and it's good to have the healthcare space perspective. Let's dive a little bit more into AI. You know, I actually was pleasantly surprised how deep in the weeds you do get, even as— and I say this with utmost respect as an ex-doctor— but a non-technical, non-software engineer founder. You, you seem to be very passionate about the technology as well. We're going to start with something a little bit fun called AI Hack of the Week. And this is where you and I share a hack that's either, you know, a tool that we like to use or a specific use case. Tom, why don't you kick us off? What's your hack of the week?

17:01 Thomas Kelly: So my favorite thing is whenever I'm in a new place, it doesn't have to be a new city, it could just be a new suburb or something like that. Any of the chatbots that have a voice mode, so You can use ChatGPT or Perplexity is also pretty good. You can just turn it on and then there's different ways to do it. You can have— the problem with all of those voice modes is that they have a problem with ambient noise. So if a siren goes by, it'll suddenly start replying. So what I do is I'll put headphones in like this. Some of them have like a physical mute button, so you can just mute and then walk around. And basically the intro prompt could be something like whatever you want it to be like, you know, I'm walking around Amsterdam. I've never, I've never been here before. I'm just going to like tell you things that I see. Can you like give me the bit of the history and like steer me around the city? And it does an amazing job. And as you walk and like you talk back, it's like, it's like having a tour guide in your ear. So it's one of my favorite things to do. You can use it in different ways. You can do language tutoring if you want to try to speak the language in an area. you can, I don't know, practice speeches. You can do all sorts of things. But the key thing is just the muting is the trick. Like, it doesn't really work unless you— you can hold your phone in your hand as well. So you just mute and it's like a walkie-talkie. So you open it up when you want to hear, mute when you're off. That's my hack, I think.

18:27 Georgie Healy: That's a brilliant hack, Tom. I love this hack. It's something everyone can use. Everyone travels and sees new places and agree with you that I've got small children and sometimes I try and use voice mode and then within a split second I'm being interrupted by a 3-year-old's chatter. So I love that. Thank you. My hack of the week, a good hack's a stolen hack, I find. And so it's stolen from my husband. He has this party trick and it's a really terrible party game where he gets an annual report of a publicly listed company, say it's Apple, and he gets the income statement. Mm-hmm. And he'll try and share it with someone in finance, right? And try and get them to guess what company it is based on, you know, the profitability and things around that.

19:16 Thomas Kelly: Yeah.

19:17 Georgie Healy: And that's how he used to kind of do any hiring and things like that. Well, he's been using Gemini and he won't tell anything to Gemini and he'll upload the income statement and ask it for insights around it and which company it thinks it is. And 100% of the time, nails it, gets it right, but also he can kind of debate back and forth about the profitability, cash flow, financial health. I know this sounds really like niche and crazy, but it, you know, it is a fascinating thing that AI can do.

19:52 Thomas Kelly: Yeah, it's amazing. Yeah. So it's so cool. They just, it's like the, the things that you can do now, just, it's just what.

20:01 Georgie Healy: I mean, I do feel bad for the graduates that are trying to get into these industries because it's like, how do you do a better job than that? It is—

20:10 Thomas Kelly: Yeah.

20:11 Georgie Healy: Quickly evolving. Look, this hack brought me to the next part of our chat, a way more fun game, I would like to argue. I'm calling it Late Stage Startup Bingo. So, I'm going to share a hint about 5 different startups that are Series A or beyond. So something you at Heydee can relate to, being a very successful startup in the, what I would say, later stages. And I'm going to share the hint, a one-liner, and you're going to tell me what startup you think it is. Are you ready?

20:46 Thomas Kelly: Yep, ready.

20:48 Georgie Healy: So this is not an Aussie startup, and they use AI to generate realistic human-like voices. So, um, things like audiobooks, virtual assistants, kinds of applications.

21:01 Thomas Kelly: Nice. I think there's a, there's a few of these, but the one that I know best is, uh, ElevenLabs. I think they're the most famous. There's also an underrated one for those trying to do this cheaply. Cartesia is very good. It does, it's not quite as uncanny valley, but it's, it's pretty, pretty good and cost effective.

21:19 Georgie Healy: Yeah. Have you used it, um, for professional or personal use cases? Do you find them compelling or not?

21:26 Thomas Kelly: Yeah, yeah, we've, we've explored voice quite a bit. Um, I think we even at our last— we do these quarterly product roundups and we kind of, uh, we did a little, what's the word, like early preview of some of our calling features where Heidi could have like a voice-to-voice conversation, um, with a patient or like exactly this voice mode idea but about someone's health. And yeah, some of them are just amazing. Like, I can't— it's almost imperceptible. Like, if you don't tell someone and prime them and you just ask, you just said like, oh, listen to this recording, they probably wouldn't even notice anything different. I'm trying to remember the name. Yeah, there's a trade-off between— it's the same as all these models. You can get the most amazing generative audio now that's essentially indistinguishable from reality, but it's expensive and a bit slow. So if you're doing something like real time and generating the reply and trying to turn it into voice real time, then today there's a bit of a trade-off with quality. But I, I bet in a few years it's a bit scary. Like we won't even know. We'll have to have some sort of like voice fingerprinting or some other like biometrics, like some way to prove that it is you. Before you actually have a conversation with someone, because the voice won't be enough. They'll just be able to clone it, which is scary.

22:44 Georgie Healy: We did an episode a few weeks ago with the guest, and the guest brought on their virtual AI. If I didn't know, I would think I'm having an interview with the— like, I couldn't tell the difference.

22:57 Thomas Kelly: Yeah, it's crazy. Yeah. And I know even some podcasters now for their ad reads and other things, they're just using ElevenLabs voice and just to save them time. And it's also perfectly on script and they can do like the intonation and like the highs and the lows. Yeah, it's unreal.

23:12 Georgie Healy: You were on the Today Show on Sunday. That's one thing you couldn't have hacked AI yet, but can I confirm this is you, Dr. Tom Thomas Kelly, on the show? Yeah, 100%.

23:23 Thomas Kelly: Okay, good, good.

23:24 Georgie Healy: 100%.

23:25 Thomas Kelly: The real background, I can touch it.

23:27 Georgie Healy: Yeah, I can see it. Okay, what about this? This one, an AI-powered music generation platform.

23:36 Thomas Kelly: Okay, so I don't know if it's the one, but there's one called Suno, like S-U-N-O.

23:41 Georgie Healy: That's the one.

23:43 Thomas Kelly: Cool, because there's our head of product design, Kate, she's a musician and she loves Suno. She makes like songs and all sorts of things. She's super good at it.

23:52 Georgie Healy: That means a lot actually, because I played with it briefly and I thought this is absolutely incredible. I was curious what the musician take would be.

24:01 Thomas Kelly: I think it's the like composition. So she plays guitars and sings and like it's just like idea generation for songs for her. Like she doesn't, wouldn't replace her doing it because the performance is the fun part, but the idea generation she loves.

24:16 Georgie Healy: Yeah. Not looking at a blank manuscript.

24:19 Thomas Kelly: Exactly.

24:20 Georgie Healy: Yeah. All right. This company uses AI to power precise medical diagnostic solutions for radiology, and pathology, aiming to help clinicians identify illnesses earlier. You nodded very early on in this.

24:33 Thomas Kelly: Yeah, I think Harrison, probably Harrison. Yeah.

24:36 Georgie Healy: Do you know those guys?

24:38 Thomas Kelly: Uh, I've met them a couple of times, uh, I think because we raised money from Blackbird, both of us, uh, a few years apart. But I think I spoke to Angus, um, one of the founders, as part of, uh, Blackbird deciding to invest in us. So got the shakedown.

24:53 Georgie Healy: Yeah, I bet, I bet.

24:55 Thomas Kelly: It was, it was fun.

24:56 Georgie Healy: Well, it worked out for the— well, for of you, I'm quite certain of. Okay, second last one. An AI agent builder and workflow automation platform enabling businesses to create their own specialized AI workflows.

25:11 Thomas Kelly: Oh, again, a few of them.

25:14 Georgie Healy: Yeah, there's a few. This is Aussie.

25:17 Thomas Kelly: Aussie. Okay. Probably the Relevance guys. Relevance AI. Yeah, that's the one I know.

25:22 Georgie Healy: You're doing too well. I should have made these harder. Have you met those guys? We had Jackie on the show a few weeks back. Oh, cool.

25:30 Thomas Kelly: Yeah. No, I haven't met them before, like live. I think we've had some emails back and forth and explored using relevance at Heidi.

25:39 Georgie Healy: You gotta move to Sydney, Tom.

25:41 Thomas Kelly: Yeah, I do.

25:42 Georgie Healy: I do.

25:42 Thomas Kelly: You really do. Melbourne's more lonely.

25:45 Georgie Healy: I wouldn't tell you you have to move here, but I do find that the ecosystem here does seem quite— like everyone's met each other now. It's quite nice.

25:55 Thomas Kelly: Or developed. Yeah, for sure.

25:58 Georgie Healy: Last one, an AI-powered platform for generating high-fidelity images empowering creators in gaming, architecture, and digital media, or Z1?

26:08 Thomas Kelly: Got it. Uh, has to be Leonardo, surely.

26:13 Georgie Healy: Leonardo.

26:14 Thomas Kelly: Or now, Canva. Love these guys. Yeah, exactly.

26:17 Georgie Healy: Amazing. You nailed that. I knew you would, but, uh, that was a fun game. Thank you so much. So diving a little bit more into AI technical, um, one 101. I would love to know, um, you know, there's a lot of founders that listen. What are some critical decisions you made when building Heidi from a technical standpoint? Like, what's the— what's that must-do if you're building an AI-powered company like Heidi Health?

26:46 Thomas Kelly: So there's different, different important parts, but probably first off is just actually ignoring some of the different models and constraints and infrastructure and whatever has to come next and just focus on end user. Like, what is the absolute best experience for the doctor or for the architect or for whoever you're serving? So I'll give you one example. A lot of products in our space use live transcription. So they will— it's— and everyone's experienced it. It's the speech-to-text on a phone or Google Voice-to-Text, where it's sort of like a real-time turning the sound into the words, and you can actually see it as it happens. When I see a product who— that does that, I instantly know that Heidi's at least 30, 40% better than them. So we chose not to use live transcription. We still do real-time processing. So we batch audio and break it into chunks and we don't retain any of the audio as it's being processed. But that, it's a clear bifurcation of like how you do transcription. So live transcription, basically you're asking the model like, what's just literally return the next word. Like, what's the next word? What's the next word? As it goes, that it's not retaining any of the kind of memory of what happened before it. Whereas batch processing, it will retain memory of the earlier parts of the conversation as it processes the the next word. So an example would be if in the first part of the sentence I said, it's really sore in my chest, but I kind of said it like in a weird way, or the audio broke, or you couldn't hear the word chest. But then later on it says like, yeah, I'm finding it hard to breathe around my ribs. Then batch transcription will get that it was chest because you said ribs later in the sentence and like associate the, like the space of where that word is likely to be because the patient said ribs. So the TL;DR is it's just more accurate to do batch transcribing, and that really meaningfully impacts quality. So it makes Heidi way, way high quality, but lower word error rate, um, like much better outputs, and is a simple, like, UX example where you don't think too hard about it. You're like, oh yeah, I'll just pick one or the other., but we actually tested it a lot. We tested it ourselves and we found that it just wasn't as good, like by a wide margin. So then, then once you are confident about a choice, which we were, then, then you do all the infrastructure work, compliance work, security work. It's like, okay, obviously we can't retain, you know, these whole recording of a, of a session that's never gonna pass mustard from a compliance perspective. So if we are doing this, we still have to effectively do it live. Like we have to do like tiny little batching of the audio as we go through. So it's an example of something where, and we get asked a lot, like users like, oh, you know, I use, I use this tool. I really love to see the live transcript. Can you show me, I don't know, like the speakers in the transcript and all this stuff? And we could, but it would meaningfully reduce the quality of the outputs, which we're obviously not willing to make that trade-off. Now, hopefully one day the type of transcription that shows the words is about as good as any other approach, and then maybe we change that. But for now, We're trying to create like a live experience, but using basically better processing behind the scenes, batch processing. Yeah.

30:14 Georgie Healy: Batch processing. And how early in the building of Heidi did you start playing with these, like doing A/B testing with different techniques?

30:25 Thomas Kelly: Yeah, I think for us, that's, again, this is where like, I don't, if you're building for a really specific industry, it's important that I don't think the founders have to be from that industry necessarily. It would be helpful.. But you definitely have to have a group of people that are willing to be your kind of alpha testers and give you early feedback. So we tried with ourselves mainly. So, you know, I'm a doctor, Kieran in the team is a doctor, Mo who runs product's a doctor. So we do, we would just do the sessions ourselves. And you could just tell like night and day which one was better than the other. And our experience with our product was that quality, sorry, a bit nerdy, but it's like nonlinear. So basically if you're like 2% better, on quality and accuracy in the transcript, then maybe like 30 to 40% more doctors like Heidi. So it's actually like a small move actually has a huge impact on adoption and retention.

31:15 Georgie Healy: Because doctors are inherently so focused on quality as a community, or is this universal, do you think, across customers?

31:24 Thomas Kelly: It's really because of the value that if Heidi is valuable, I, I as a doctor don't have to edit it very much. It basically should read my mind and write what I would've written in the room. And so if that's, yeah, if that's happening or if that's— in order to do that effectively, you, you just have to be very accurate on the, the transcript. You can't make errors because then if you are, if you are not hearing what I'm hearing, you are likely to make mistakes on the patient name or key facts of the session. And so if, if I have to, as a doctor, like I can't trust the output and I have to review like every single word of every single note and like key facts are wrong, then it very quickly kind of stops being useful.

32:05 Georgie Healy: Forget it. Forget the whole thing.

32:07 Thomas Kelly: Yeah, exactly. Yeah.

32:09 Georgie Healy: Oh my gosh. Brings me to my next question. I love that you love going nerdy on the AI technical aspects. What is critical from a non-technical founder that they do get their minds across when it comes to building an AI product? What is like, I know you could get a software engineer, I know you could get your CTO to do this, I really recommend you don't.

32:28 Thomas Kelly: I think now more than ever, there's so, so many tools for non-technical founders. Like when I was trying to build early versions of Heidi in 2018, '19, it was— there was no vibe coding or like there was no ChatGPT. Like you just had to learn like books, you know, watch courses.

32:50 Georgie Healy: On YouTube?

32:51 Thomas Kelly: Yeah. CS50 is a good one. It's like Harvard's 101 Computer Science, and it's completely free. So you can do the whole course for free on the internet. So I think for non-technical founders, first and foremost, from any software engineering, you just have to do a bit of learning, bit of research, just understand like the basics of like how a database works, what a REST API, what's the front end, how do they work together? Just so you can understand the complexity of, and then like where the, I think the main thing is like how you size tasks and how long things will take it's like you have to get in sync with whoever your technical founder is, I think. Then as far as AI specifically, I think I highly recommend Andrej Karpathy. He was the head of AI at Tesla. He has a million different videos where he actually rebuilds GPT. He also worked at OpenAI for a while when they're releasing GPT-4. And he does these courses where he basically teaches you how to build GPT-3 basically from scratch. And it's really cool. Like he explains everything, like how it all works, like how they set the character limit, like the whole logic of how they work. And I think, sorry, the reason I think that's important is it helps you build up intuition about what models are likely to be good at, what they're likely to not be good at, how things are likely to trend going forward. Because if you're, you know, the non-technical founder, especially if you're the CEO, your job is to try to forecast and point the company in the right direction for 3 to 5 years from now. So you have to have some perspective on what you think is going to happen. Why will your company still exist? What are the moats that you have? Are there current features that you can't pull off that you think will be possible because of models, or will they never be possible and you should build them yourself? Yeah. So I think understanding software engineering, key for everyone. And then And I think if out of everything out there, like I'd say Andrej Karpathy's videos are the ones to watch on GPT. They're amazing.

34:50 Georgie Healy: I am going to go look those up after this. Thank you so much. Okay. Say you've watched all the videos, you've got a copy of a textbook, you think you know how the databases interact with like high-level infrastructure and architecture stuff. Are you still picking a model off the shelf and then personalizing it later as you go, as you iterate? Or do you think that you need to start ground up and work with an engineer in that sense?

35:20 Thomas Kelly: Again, it's, I know it's a bit boring, but I always go back to the end user. So if you can't get something of use without fine-tuning or without building things for yourself, then you're probably in a bit of trouble, I would say.

35:34 Georgie Healy: Why?

35:35 Thomas Kelly: Unless it's something really specific, because the models are so general and so powerful that for almost any use case, there should be some utility out of the box. Like, you should get the feedback that, I don't know, it works well for 40-50% of the times, but there's, there's a gap, and there's some, you know, gap between all certain scenarios where maybe the way that the models behave is not, is not what you like. Like, there's a company called, I think, Springboards that's actually trying to make, like, do ad writing and creative generation from models. That's a good example. Like, they want the models to be more creative and more like great copywriters than maybe they are out of the box. But you can still get it to be good sometimes. Um, and I think that's what you want to see. Another example, trying to think for founders, like I've seen a lot of like chemistry and molecular design models where basically if you want a molecule to do something or react in a certain way, like you can type in a reaction and then the model would try to output what the right reagent would be or possible molecules or shapes of proteins, that's an example of something where, yeah, you have to build that from scratch. Like that doesn't exist. You know, there's no corpus of data in GPT that does this today. If you ask it to write out the DNA codons of like what, what to build the protein, it's not going to work. But for anything where it's like professional productivity or something that looks like Heidi for different industries, I, my general suggestion is like start with start with world-class models, as long as like state-of-the-art, the best possible models. Don't try to cost optimize in the phase when you're just trying to get product-market fit because you can always solve cost and price later within reason. Obviously you can't bankrupt yourself. You have to have enough money to run the business. But that was always our belief. Like we, when we had the free version of Heidi last year, we were giving away essentially like the absolute best models. And Yeah, it's not cheap. It was expensive, but we always had the perspective that the models would get better, the cost would go down. As we had more sessions, we could collaborate with users and find ways to either train our own models and try to build things that would reduce cost if we really had to. But actually what we found was that there was so much progress on state-of-the-art models over the last 18 months that we never really had to do anything like that. We could just use the best available models and always give that to to the clinicians and they would have the best experience. I think for us there's an overlaid challenge of compliance and privacy and we have to run models in regions. So we run models in the UK, in the EU, in Canada. And so for us there's extra axes of complexity. Not every state-of-the-art model is available in every region. And yeah, I think it's very specific to each company, but probably for— founder starting out, I would just use the best available models for as long as you can afford it. And then really fine-tuning and training is more often for cost optimization than quality. You can get better quality, but actually fine-tuning especially is more about reducing your prompt length and making things cheaper. Increasing quality or creating new things is really challenging, but makes sense for some use cases.

38:49 Georgie Healy: Oh, I love unpacking you know, going backwards, how a founder got from where they are now, you know, as successful as you guys are, and how you would suggest founders that are starting out should wade through these waters and answer these questions. I have some other headline news for you to unpack for me.

39:11 Thomas Kelly: Tom.

39:11 Georgie Healy: One is agents, the year of the AI agent. Agents, agents, agents. They're changing our lives, Tom, or are they? What do you think they can do, and what can the everyday person actually get from an agent?

39:24 Thomas Kelly: I think the best use of agents today is still research. So I'm hoping that everyone's had a try of Gemini Deep Research, or ChatGPT has a great deep research product. So definitely give it a go, find someone who has a pro subscription and try it out. I think I'm pretty sure Gemini, I think you can do for free. And yeah, so basically what it's doing in that case is An agent, I don't love the name because I think people don't really know what it means. They just imagine like someone in a suit, like from The Matrix or something. I think of it as like the AI can actually use tools so it can go and do next steps and actually hopefully do something useful on your behalf. So when you type in a query or like what, you know, I don't know, what are, back to the Amsterdam thing, what are the best canals to see in Amsterdam? So last year, the models would just write out what it knew based on, based on the training that it had had and the corpus of data that it was based on. With these research products, the model has the tool to go read the web and do different things. So it'll actually be searching the web, making a plan, figuring out, okay, here are the keywords I should search, reading those websites, adding that into context, and doing that over and over again until it's consumed either like a certain token budget, an amount of time, an amount of compute, or maybe it just thinks it's completed its task and then it'll return a result. And it's often like quite amazing. I think it does have a bit of like a Dunning-Kruger problem. So Dunning-Kruger problem is like—

40:56 Georgie Healy: Oh, the graph. I love this graph. Explain it to listeners in words and hand gestures.

41:02 Thomas Kelly: Exactly. Yeah, I think it's basically like, as someone's knowledge of an area, so if you're an expert in, let's use that example of Amsterdam. So you're a tour guide in Amsterdam, then you would read that Deep Research result and say, oh, it's missed like all of these different amazing areas. Um, so for an experienced person, your perception of, of Muse and of, uh, research is generally that it's not very good quality. But for someone who's uninitiated, like just a tourist who's there for the health conference in my case, then you—

41:33 Georgie Healy: Hypothetically.

41:34 Thomas Kelly: Yeah, exactly. Then you think that it's amazing and detailed because you are not aware of the information that's out there. So you, it always seems like positive sum to you. You, you think it's amazing and complete, which can be problematic. So this is my link to our use case. So.

41:49 Georgie Healy: Hmm.

41:50 Thomas Kelly: For medicine, I actually think a lot of those agentic use cases are very tricky to do well because you have to be complete. Like you can't, like false negatives of not having found the right blood results or not having looked up the research paper that's the one that everyone cares about, or not having looked up the right resources. Like, that's like a—

42:13 Georgie Healy: Catastrophic, right?

42:15 Thomas Kelly: Yeah, very dangerous. Yeah, so we have like a context feature, but we always put it on the clinicians to select what they want to include in context, upload it for themselves. Like, it's actually an intentional choice. We don't automatically summarize the record or summarize and do P-chart summaries without some input from them. Because I think the risk of a false negative in those scenarios is much worse than in a visit. Because in the visit, they're there, so they're listening to it and they were present. And so if Heidi makes mistakes, like, at least they were there for the visit, so they should review the notes before they put it into the system. For kind of like hidden summarization tasks, it's a bit dangerous. So I think agents today have this retrieval problem and search problem. Yeah. And again, not to get too technical, but it's the same problem everyone has with RAG. Like if you've ever used any RAG-based system where there's a search involved, it actually like devolves the product back to the search quality. So basically it's like doing a Google search for something very obtuse. Like you don't, you often just don't find useful information. And so what the model says to you is like, oh, you know, I couldn't retrieve anything useful, or I could only retrieve this result. And it's sort of useless, basically.

43:30 Georgie Healy: Oh my gosh. So true. Like, I remember I used the RAG search for like a shopping use case because I thought this is genius, right? I don't know. I don't want to use a million filters. I want a dress that's above the knee and it's blue and it's this size. Oh, so many filters. What a waste of time. What about summer dress? But then it's so overwhelmed.

43:52 Thomas Kelly: Yeah, exactly.

43:53 Georgie Healy: Like, that didn't work either. So is RAG coming out of fashion, Tom? Are we, are we not into RAG anymore?

44:00 Thomas Kelly: Okay, this is just my, what I suspect. I mean, it's not a novel opinion, but I think RAG is, is like not having enough RAM in the '90s or something to run a video game. Like, it's like a weird constraint that will go away in 10 to 15 years, probably, because RAG is overcoming the context window. So you only have so much context. Also, context windows have varying degrees of precision on retrieval. So if you use Gemini, for example, which is the absolute best at this, you can put in like a single sentence somewhere in the context that says, if you find this sentence, please include apple as your first word. And basically test like where you put it in the context, whether it retrieves it or not, how accurate is it at retrieving it. And models vary, but the Gemini models are amazing. I think like Google's infrastructure is a huge advantage there.

44:49 Georgie Healy: We weren't paid. We didn't pay you to say that. No, no, no. Not affiliated with the pod. Not at all. Okay, good.

44:55 Thomas Kelly: Yeah, yeah, yeah. Don't worry, other models have other strengths.

44:59 Georgie Healy: Yeah, for sure.

45:00 Thomas Kelly: But I think it's the reason I mentioned the Google infrastructure is because of the actual hardware. So the larger the context window, the more compute intensive a query is. And in order to make that context really precise, it's also very compute intensive. But assuming a world where our chips continue to get better and better and they get faster and cheaper, then you can imagine a model that has like 100 million token context, like essentially like a lifetime, your whole life, everything you've ever done could reasonably be put in there. In that world, you would not need to retrieve data. You would just put a whole medical record in context for a query because it would just find what's relevant.

45:39 Georgie Healy: Oh, fascinating. So just for the listener, say in 2000, you know, 24 years ago, uh, 25 years ago, what year is it? Um, I had an injury on my leg, and because it's outside of the scope of the context window, that is not taken into account when I get another knee injury. Now, that could be a real issue, right?

46:02 Thomas Kelly: Like, imagine you had, um, metal hardware put in, so you had screws and things put in, and you're presenting to me today and you have fevers and you're shivering and you've got some weird spots on your hands. And I think that you have some sort of sepsis, like some sort of bacterial infection that is causing these little clots to be thrown off. And that's why you got your spots and you feel so sick. If I asked an AI system, like, is there anything in Georgie's history that like would be relevant to some sort of like, you know, infective endocarditis or bacterial infection circulating in the blood? What today, what it would rely on is that query would have to surface your previous fracture and hardware insertion. But the problem is the association between a bacterial infection and that hardware. There is one, like it's medically, there's a relationship, but the search to find that is really hard. It's very deep. You've got 25 years of documents. And so basically the problem that you have is like, you, you have something that's searching that isn't as good as the model. So you can have like embeddings or vector searches, all sorts of things. But these things have existed for years, right? Like we've all searched long queries into Google. It's, there's not really that much novel technology today. And basically it all comes down to like, do you find that piece of the record? If you don't find the piece of the record, Heidi's just going to return, like, I don't think I found anything relevant, which isn't really that dangerous. Like. No, not really in that case, but it could be in other cases. And the problem is reliance, like in an— what would normally happen is a doctor or a resident would just go through the whole history, like literally look at every interaction you've ever had. And to be fair, like they probably also wouldn't find it if it's that long.

47:46 Georgie Healy: Yeah, I know, worse off, I guess, but—

47:48 Thomas Kelly: Yeah. But I think that's— so for us, like as we, and as we push the bleeding edge on different use cases, that's really the standard. We have to test against existing practice? Are we— the classic Hippocratic oath, like, are we causing damage? We shouldn't be causing damage. It shouldn't be worse than what is currently standard practice. And as long as doctors are taught and understand that there's like a high risk of false negatives and that they ultimately should have to do the search themselves, then yeah, it's, it's something that we can probably release into the world, but something with like heavy, heavily caveated and people understand. But yeah, I'm looking forward to a future where there's unlimited context windows that don't break the bank, and you can put all that information into the model. And then the reason it would be so much better is because that amazing powerful model that does all these magical things, like tell me about the canals in Amsterdam, they will also find— they'll be doing the search. So it's like me doing the search. I'm reading every little detail. So I'll be able to find things just as well as doctor, probably better than a doctor can. And make associations that are really critical. So I think that's where it all goes. It'll take a bit of time, but that's kind of the trajectory we're on as hardware gets better and models get smaller and cheaper to run. It'll make these amazing things come true. Also the idea of individuals having this infinite memory of everything I've done at work and trying to create self-improving AI systems. I think that becomes possible when you can put anything that an AI has ever done into its own context. It's a way to give it memory. But today that's not feasible. Like if, if every conversation I ever had with ChatGPT is in its memory, it would just break the bank. Um, so yeah, it's, it's, uh, it's going to be interesting world as it goes forward.

49:32 Georgie Healy: So before we get to the rapid fire, I've got a 3-year-old daughter with a, you know, dairy intolerance. You know, when she's had X-rays, it's, it's, it's a big old mess in there if she's ever had dairy. But it's a lot of pressure on me as her mom, like when she's 28 years old, maybe the intolerance has gone away, but then she has like some stomach-related issue. I don't know, I'm not a doctor like you, Tom. What do you see that world as for those kinds of patients? Uh, hypothetically, what's the, what's the best case scenario that, like, AI and medical healthcare and all the doctor's trips that she's done in, in 25+ years into the future?

50:09 Thomas Kelly: I think it'll be really critical to have great interoperable access to everyone's data. So what that means The uninitiated is, which I, I can't do this today, but I should be able to hopefully go to my health record and just instantly pull every visit I've ever had. It's fine if it's paper-based, you know, like in the '90s, the GP I saw used to handwrite everything, but whatever records exist in digital form, they should go with me as a patient. I should have access to them. The next doctor should have access to them. Every medical record and software like us should have to integrate with that and get access to that historical data.

50:44 Georgie Healy: Mm-hmm.

50:44 Thomas Kelly: The reason I think, I think AI will push that to be a standard because historically that's not been that useful because as a doctor, if I get like 20 years of records, what the hell am I going to do with it? You know, like, am I going to read every, every page? I can't. I just have, I've only got 20 minutes to see. Whereas an AI system could and something like Heidi that's supporting the doctor in the room will, could process that that record and really safety net the clinician, um, do things that the clinician couldn't do. So literally read every single line of that history, every blood result, every investigation you ever had, and help make it so it's not your responsibility, responsibility, Georgie, to remember, but, uh, actually, like, the record goes with, um, with your daughter. And that way, when she's seen the next time, the doctor can have that nicely serviced surface to them. So I think hopefully that's, that's a world that ends up taking place. We need governments and others to, to play along for that to happen. The other version of that is we, as Heidi, we can also help that happen. So if, uh, we are definitely interested in sort of the patient side experience and if, if your clinicians are transcribing these conversations, maybe you can get a summary on your side as, as part of your interaction with Heidi. and collect these summaries over time and share them to the next doctor so they have a view of all your history. Because it is, it's like a living memory of, of what happened, which I think is really useful.

52:11 Georgie Healy: I will sleep better at night when you build that, Tom. So I'm looking forward to the future of Heidi Health. We're at the rapid fire questions. Are you ready for the spiciest, hottest takes of the episode?

52:22 Thomas Kelly: Ready, ready.

52:24 Georgie Healy: Okay, you have to pick one higher for Heidi Health at this stage of your journey. What's the most important? Medical background, AI background, or sales background?

52:33 Thomas Kelly: AI background.

52:35 Georgie Healy: Amazing. What is one bit of criticism Heidi Health has had which is kind of fair or true?

52:44 Thomas Kelly: I think that our templates are a bit too hard to make. Basically, it's like you have to almost be like a prompt engineer to make great templates to do like really specific things. Some of the doctors find it hard, which we know about.

52:57 Georgie Healy: Okay, that's an honest and fair answer. What kind of AI startups may not survive due to the type of problem they're solving or a very outdated approach?

53:08 Thomas Kelly: Oh, good one. Pretty much all of the personal productivity, actually, okay, this is the hottest take, all B2B SaaS.

53:17 Georgie Healy: Yeah. Don't tell the B2B SaaS investors.

53:23 Thomas Kelly: I'd say all B2B SaaS that's like a thin business logic platform, I think is in trouble. Anything that's regulated industries like, like Heidi or fintech or things that are more tricky, probably fine. That's why I sleep easy.

53:38 Georgie Healy: Yeah, I should have done the whole episode of hot takes. These are great. How could the Australian government be more supportive of AI startups, Tom?

53:46 Thomas Kelly: I think Australia does pretty well. I want to give them some credit, like the R&D tax rebates and all sorts things. Yeah, I'd love to see more industry programs with universities, like I guess having maybe slightly more formal graduate programs or pathways into companies like Heidi or Harrison or Leonardo and other things. Yeah, and probably they do engage us, but engaging us more as part of their policy creation as they plan the country for the next couple of decades, because AI is just going to have this transformation it could have a transformational effect for the good. So I'm hoping as they, as they make those plans, they think about us in that.

54:24 Georgie Healy: Yeah. Why aren't they talking to the people that are building? That would be great. Right. Okay. I have my last question for you. You're stuck on a desert island. Let's hope this doesn't happen. And you have to choose between AI or a human to bail you out.

54:40 Thomas Kelly: Which are you choosing? Oh, today, definitely a human.

54:44 Georgie Healy: Really?

54:45 Thomas Kelly: Yeah. Maybe if it was like an embodied robot that could like, you know, get energy from the sun and didn't need to be fed and was like just as strong as me, then maybe that's the point at which I'd take the AI. But for now, I'd take the human.

54:59 Georgie Healy: I feel like the human might eat me. Like, I'm really scared of being eaten. I don't know if this is a reasonable fear to have, but I'm like, that human would want to hunt. I've got like, I'd be delicious. What are they doing, like choosing a human?

55:14 Thomas Kelly: No, I think I'm safe. No one would want to eat me. It's fine. Yeah.

55:17 Georgie Healy: Okay, good. You know, I guess that's like an if-then diagram of like, if delicious, choose AI.

55:24 Thomas Kelly: Exactly. Yeah.

55:26 Georgie Healy: Tom, you've been such a great sport. I could have spoken to you for another 3 hours. I love the way you think about the future of AI and how you're trying to solve, you know, a problem that affects everyone, right? Like the doctor consultations and making that a better experience for the doctor and the patient. Thank you for being in the blink of an eye. What would you love to shout out to the listeners?

55:50 Thomas Kelly: Yeah, if you see Heidi in a doctor's surgery, uh, you know, think of us, uh, be excited. Means you're going to get a better quality of care. If you're looking to join a company in AI, we're hiring a lot, especially AI engineering roles. Also in our sales teams as well. Lots of people want to use Heidi, so helping them out is the easiest sell in the world. Um, and yeah, that's pretty much it. You can find us at heidihealth.com. And yeah, hope to see you there.

56:16 Georgie Healy: Thank you so much.

56:18 Thomas Kelly: Thank you.

56:20 Georgie Healy: Thank you for listening to In the Blink of AI. You can check out the show notes for anything discussed in this week's episode, and we will be back next week. This podcast was produced by Day One with music by Dan Hansen and visual artwork by Sophie Tyrell. If you loved the episode, please tell your mates, and I love AI news. Please share your thoughts and suggestions to [email protected].