Bản chép lời đầy đủ của #560: Building a Research OS: From Django to 30,000 Samples

primary area of practice is in the field of inflammatory bowel disease that comprises 00:04:02.960 --> 00:04:09.280 ulcerative colitis and Crohn's disease as the kind of conditions that we deal with when 00:04:09.280 --> 00:04:10.480 we see patients in clinic. 00:04:10.740 --> 00:04:15.539 But in parallel, I also work in research, trying to find out the causes of these conditions 00:04:16.019 --> 00:04:20.220 together with the GUT Translational Research Group here at the University of Glasgow. 00:04:20.700 --> 00:04:25.400 Personally, for this podcast, we use Python a lot in our work, both for research, 00:04:25.760 --> 00:04:29.640 so trying to analyze the data that we're doing, but also as well as for the infrastructure that 00:04:29.670 --> 00:04:33.540 we're running. So how do we actually deliver the research studies that we have to deliver? 00:04:34.300 --> 00:04:39.680 And personally, as a researcher, I've also got an interest in agentic AI and machine learning, 00:04:39.730 --> 00:04:43.739 and that's what attracted me to try to use some of these techniques to see if we can find 00:04:43.980 --> 00:04:48.700 new breakthroughs that can bring, you know, new treatments for our patients that we see in clinic 00:04:48.960 --> 00:04:53.460 every day. Awesome. Very cool research. That's the kind of stuff that can make a big difference 00:04:53.680 --> 00:04:58.040 for people's lives. If you can make them better, right? It's like, I couldn't leave the house, 00:04:58.260 --> 00:05:02.780 but now I'm back with my friends or whatever, right? Yeah. I mean, you know, the work that we 00:05:02.960 --> 00:05:09.139 do is motivated in the patients that we see day to day. Inflammatory bowel disease is not very 00:05:09.160 --> 00:05:15.140 common. It is rare, but it is growing not just in developed countries, but also in the developing 00:05:15.440 --> 00:05:20.200 world. And we see a great burden of this disease coming the next 10 to 20 years. What do you think? 00:05:20.480 --> 00:05:26.000 Diet? Environment? Why is it growing? Obviously, that's the research a little bit, right? But 00:05:26.320 --> 00:05:30.840 what is your ideas here? Well, these are really complex conditions, and the environment is 00:05:31.020 --> 00:05:36.340 absolutely a major player in this. The diet that we're eating nowadays is different from what we 00:05:36.280 --> 00:05:41.000 used to eat. But when we dive into the science of it, it's a lot more complex than just that. 00:05:41.280 --> 00:05:46.120 There are some patients who have genetic susceptibility, although not all. And we know 00:05:46.120 --> 00:05:51.420 that there is this problem that happens when the immune system reacts to our environment. And for 00:05:51.560 --> 00:05:55.940 some reason, the alarms don't switch off and the inflammation keeps happening, which leads to the 00:05:56.300 --> 00:06:03.379 symptoms and the problems that we see every day in clinic. So let's start by examining all the 00:06:03.400 -->

00:06:09.340 the research and the work you did and maybe the 2020 version of your research projects. 00:06:09.680 --> 00:06:10.820 It all started with Django, right? 00:06:11.260 --> 00:06:11.460 Yes. 00:06:12.300 --> 00:06:15.820 So back in 2020, I was still training as a gastroenterologist. 00:06:16.140 --> 00:06:20.320 So over here, we do a period of registrar training, we call it. 00:06:20.880 --> 00:06:27.180 And as part of my developing interest in understanding these medical conditions, I sign up to do a 00:06:27.300 --> 00:06:30.660 research study with one of the principal investigators. 00:06:31.280 --> 00:06:37.620 And the proposed study was basically to recruit quite a few patients from three different cities, 00:06:38.120 --> 00:06:43.660 multiple hospitals, and collect extra samples that we can run science experiments on to work 00:06:43.780 --> 00:06:51.160 out what's going on with an immune system. So when we first started, that was 2020. And I know 00:06:51.600 --> 00:06:57.379 we all remember that in 2020, you know, we had COVID and COVID had a significant impact both on 00:06:57.400 --> 00:07:02.660 hospital operations, but also on research operations. And when I sat down to look at 00:07:02.830 --> 00:07:07.300 what we were supposed to do in terms of the research studies, the first challenge that I 00:07:07.500 --> 00:07:12.120 faced was calculating the number of samples we were going to do. So we were going to recruit about 00:07:12.400 --> 00:07:17.440 200 people and we're going to follow them up every three months. And at every time point, 00:07:17.550 --> 00:07:22.999 we would take extra blood samples, extra stool samples, and saliva samples. So that was a lot 00:07:23.020 --> 00:07:29.280 of extra samples and a simple back of the envelope calculation came to about 30,000 samples that we 00:07:29.280 --> 00:07:33.180 were going to generate over a couple of years. Now, when I first saw that challenge, I started 00:07:33.460 --> 00:07:37.640 asking everybody around me, how are we going to keep track of all that? How are we actually going 00:07:37.800 --> 00:07:44.760 to deliver that problem? How do we solve that problem? And from my chat with a lot of people, 00:07:45.180 --> 00:07:48.920 the common way that researchers track the samples is to use Microsoft Excel. 00:07:49.540 --> 00:07:53.180 That's exactly what I was thinking is like, how big is the Excel file? 00:07:53.400 --> 00:07:56.820 And, you know, how many little worksheet tabs does it have? 00:07:57.160 --> 00:08:01.960 Yeah, I mean, most research teams do not really have a lab system because lab systems, you 00:08:01.960 --> 00:08:06.260 know, enterprise software, which is really expensive to procure, it'll take you like 00:08:06.440 --> 00:08:07.420 six months to set it up. 00:08:07.620 --> 00:08:12.180 And it's usually designed really for hospital operations where you're taking, you know, 00:08:12.540 --> 00:08:13.280 millions of samples. 00:08:13.480 --> 00:08:18.500 And it's not really designed for a one-off research study across a lot of different spaces. 00:08:18.960 --> 00:08:23.840 So then I started looking into what are the, you know, are there off the shelf options or 00:08:24.280 --> 00:08:26.540 stuff that we could use to just get up and running? 00:08:26.980 --> 00:08:32.180 But after looking at a few different options, I think I had to bite the bullet and decided 00:08:32.479 --> 00:08:36.479 that, you know, the best thing was to go and write a web application with Django. 00:08:38.060 --> 00:08:40.400 What was your programming experience at this point? 00:08:40.840 --> 00:08:41.979 How good of a programmer were you? 00:08:42.940 --> 00:08:45.080 So I wasn't very good of a programmer. 00:08:45.140 --> 00:08:50.500 I mean, what I did was basically, you know, write some websites in high school. 00:08:51.000 --> 00:08:53.080 And back in those days, we used to use Notepad. 00:08:53.340 --> 00:08:57.200 We used to open the brackets of HTML manually by hand. 00:08:57.460 --> 00:08:58.040 It was rough. 00:08:58.240 --> 00:08:59.860 Those were tough times, I remember. 00:09:00.240 --> 00:09:03.380 Yeah, I don't know how many people listening to this podcast can relate. 00:09:03.800 --> 00:09:06.000 But we used to write, you know, the head, the body and stuff. 00:09:06.240 --> 00:09:11.020 And then later on, I did use a bit of programming to do some, you know, statistics, a bit of, 00:09:12.120 --> 00:09:15.200 writing an R script or a Python script just to generate a graph. 00:09:15.490 --> 00:09:19.180 So that's probably that level of programming experience when I was 00:09:19.560 --> 00:09:22.580 looking at how do we actually write a web app with Django. 00:09:23.440 --> 00:09:27.460 Yeah, okay. So a Django web app was kind of a 00:09:27.550 --> 00:09:31.340 pretty big stretch at that point, right? Yes, but the Django... 00:09:31.420 --> 00:09:35.260 How'd you get started? I mean, this predates all the AI, agentic stuff. 00:09:35.480 --> 00:09:39.259 You had to earn it for this one. Yeah, definitely. I mean, when I first 00:09:39.280 --> 00:09:44.580 visited a Django website. It's the web framework for perfectionists with deadlines. And I think 00:09:44.700 --> 00:09:50.140 that described exactly what a situation I was in. You know, I came out program and we had about a 00:09:50.140 --> 00:09:54.700 two year time period to deliver a study. And you can't spend forever trying to get this up and 00:09:54.840 --> 00:10:00.440 running. You just have to get the study going. So the Django tutorial was where I started. But 00:10:00.680 --> 00:10:03.960 actually, I just want to take this opportunity to thank a lot of people in the community and those 00:10:04.120 --> 00:10:09.240 listening because it's the documentation, the YouTube tutorial, your podcast. I've been listening 00:10:09.260 --> 00:10:16.000 for five years. Books that were written, Two Scoops of Django, famously. I read a ton of books

just 00:10:16.240 --> 00:10:22.420 to be able to get things up and running. Yeah, the books are great. And honestly, YouTube is really 00:10:22.440 --> 00:10:27.840 good. People knock YouTube for many reasons. You know, it's one of the ills of social media that's 00:10:27.920 --> 00:10:33.380 like scrambling their brains. But if you use YouTube for education, it's an incredible resource. 00:10:33.940 --> 00:10:37.740 Yeah. I mean, there are a lot of creators on YouTube that are just, you know, teaching lots 00:10:37.760 --> 00:10:43.240 of different things. And it's so helpful. But I also think some of the textbooks are great, you 00:10:43.260 --> 00:10:47.840 know, for things like test-driven development. You know, back in the days, I think you would only 00:10:48.140 --> 00:10:52.160 read about it in textbooks rather than on YouTube videos because it's not really an attractive topic. 00:10:53.040 --> 00:10:58.060 Yeah, you mentioned Obey the Testing Goat, which is a really fun book. And you also talked about 00:10:58.820 --> 00:11:05.660 using designing data-intensive applications. And that's by Martin Klepman. I haven't heard of that 00:11:05.640 --> 00:11:08.360 That's a pretty interesting one that people might want to check out. 00:11:08.620 --> 00:11:12.780 Yeah, down the line, as our studies matured and we started dealing with the data, I started 00:11:13.160 --> 00:11:17.800 reading some of the other data textbooks like Data Warehouse Toolkit by Kimball. 00:11:17.920 --> 00:11:23.180 I think that's the one that all data engineers used to talk about dimensional modeling and 00:11:23.200 --> 00:11:24.000 all that kind of stuff. 00:11:24.420 --> 00:11:25.220 So lots of textbooks. 00:11:25.900 --> 00:11:29.800 I think they're really helpful because they give you a kind of a holistic, fundamental 00:11:30.180 --> 00:11:35.600 approach to learning the basics and making sure you don't have gaps in what you're doing 00:11:35.620 --> 00:11:36.700 to build some of these things. 00:11:36.980 --> 00:11:39.720 When you come from the background of programming that you did, 00:11:39.880 --> 00:11:43.740 which, by the way, that's a very similar story to me as well. 00:11:43.960 --> 00:11:46.820 I studied math, and then I got involved helping people 00:11:46.960 --> 00:11:50.500 with research projects and working for a scientific research 00:11:50.620 --> 00:11:51.640 and visualization company. 00:11:52.060 --> 00:11:54.860 So you kind of build up the pieces as you go, 00:11:54.860 --> 00:11:55.460 and it's super fun. 00:11:55.880 --> 00:11:56.420 It's really neat. 00:11:56.640 --> 00:12:00.820 But there's a lot of gaps in your sort of data integrity, 00:12:01.560 --> 00:12:04.880 programming, like what is the foreign key relationship? 00:12:05.360 --> 00:12:06.040 All that. 00:12:06.300 --> 00:12:06.720 Yeah, yeah. 00:12:07.100 --> 00:12:09.520 So I think books are really important, 00:12:09.840 --> 00:12:11.720 especially for self-taught people like us.

00:12:12.040 --> 00:12:13.040 Definitely, definitely. 00:12:13.800 --> 00:12:16.140 You do learn a lot as you go. 00:12:16.660 --> 00:12:19.520 And it's quite important to be aware of what you don't know 00:12:19.740 --> 00:12:21.120 and try and close those gaps 00:12:21.800 --> 00:12:23.420 so that the applications that you write 00:12:23.560 --> 00:12:26.260 end up being reliable enough for people to depend on. 00:12:26.520 --> 00:12:29.240 Let's talk about Django for a second 00:12:29.500 --> 00:12:31.040 because you shouted out, Django, 00:12:31.540 --> 00:12:33.720 some of the features of it that kind of made that possible. 00:12:33.880 --> 00:12:37.880 I think this is one of the reasons that people choose Django, especially when they're getting started. 00:12:38.320 --> 00:12:40.400 You know, it's like it's got the built-in migrations. 00:12:40.940 --> 00:12:43.380 It's got the automatic database management, the admin. 00:12:43.860 --> 00:12:46.280 So what was it about Django that drew you to this? 00:12:46.600 --> 00:12:52.980 So I think, you know, when we looked at the landscape back in 2020, Django definitely stood out because it was a mature project. 00:12:53.380 --> 00:13:02.720 It's been around for a long time and it's batteries included, which is really important because, you know, I'm conscious that, you know, as a non-programmer coming into trying to write these things, 00:13:03.080 --> 00:13:05.560 You want your authentication framework to be battle tested. 00:13:05.980 --> 00:13:08.000 Database and migration is incredibly important. 00:13:08.050 --> 00:13:09.100 You can't lose any data. 00:13:09.480 --> 00:13:11.620 And I think just on those two features alone, 00:13:11.880 --> 00:13:13.420 we haven't even come to the admin dashboard, 00:13:13.780 --> 00:13:14.980 but just those two features alone, 00:13:15.190 --> 00:13:18.680 I think was enough to persuade me that Django is the right framework 00:13:18.920 --> 00:13:20.640 for the problem that we're dealing with. 00:13:21.540 --> 00:13:22.380 Okay, very interesting. 00:13:22.710 --> 00:13:25.840 And by the way, just big news for Django folks. 00:13:26.120 --> 00:13:27.280 Is it on their blog, maybe? 00:13:27.560 --> 00:13:30.219 Big news for Django is they just announced 00:13:30.220 --> 00:13:33.680 that they're switching the Django long-term support 00:13:34.779 --> 00:13:37.300 to yearly releases, just like Python itself, 00:13:38.059 --> 00:13:39.060 and calendar version. 00:13:39.160 --> 00:13:41.860 So there'll be Django 2028, Django 2029, 00:13:42.260 --> 00:13:44.140 and every single release is a long-term support, 00:13:44.360 --> 00:13:46.140 whereas now you're kind of juggling like, 00:13:46.260 --> 00:13:48.520 well, 5.2 is a long-term, but the one before it. 00:13:48.740 --> 00:13:50.320 I think that'll also make it a little simpler 00:13:50.440 --> 00:13:51.360 for people who are coming, 00:13:51.520 --> 00:13:53.620 like they don't have to decipher the versioning

00:13:53.900 --> 00:13:54.500 and what it means. 00:13:54.760 --> 00:13:55.720 Definitely, yep. 00:13:56.020 --> 00:13:56.800 Yeah, yeah, that'd be cool. 00:13:57.660 --> 00:13:57.960 What else? 00:13:58.800 --> 00:14:04.680 What else did you run into trying to navigate this world of becoming, running one of these 00:14:05.060 --> 00:14:11.000 like real web apps rather than just Excel or buying some off the shelf, you know, square 00:14:11.260 --> 00:14:14.400 peg round hole type of system that doesn't really fit what you're trying to do? 00:14:14.800 --> 00:14:19.120 Yeah, I mean, I think one of the steep learning curves is how do you go from a local host 00:14:19.360 --> 00:14:22.260 development project into something that's in production? 00:14:22.940 --> 00:14:25.000 Because production is obviously a whole different ballgame. 00:14:25.580 --> 00:14:29.560 you need to ensure your infrastructure in production is reliable. 00:14:30.140 --> 00:14:31.360 It's secure. It's accessible. 00:14:32.380 --> 00:14:34.960 You've got backups running, all sorts of things. 00:14:35.940 --> 00:14:39.440 And for us, one of the key things that enabled us 00:14:39.660 --> 00:14:42.760 is continuous integration and continuous deployment practices. 00:14:42.920 --> 00:14:46.300 So using GitHub Actions to deploy continuously 00:14:46.460 --> 00:14:50.620 so we can fix problems, get it into production in minutes 00:14:50.800 --> 00:14:53.860 rather than trying to do any of these things by hand manually. 00:14:54.340 --> 00:14:56.580 Yeah, the CI-CD stuff, I think, is really valuable. 00:14:56.940 --> 00:15:01.300 It's valuable for big teams, but it's also really valuable for people who are really new. 00:15:01.640 --> 00:15:07.240 Because if all I have to do is save my work to GitHub, and now that's the new stuff that's out there. 00:15:07.520 --> 00:15:09.120 That takes a lot of the complexity out. 00:15:09.260 --> 00:15:13.840 People forget just how intimidating logging into a Linux computer is. 00:15:14.120 --> 00:15:14.560 Oh, yeah. 00:15:15.140 --> 00:15:15.500 Now what? 00:15:15.860 --> 00:15:15.960 Yeah? 00:15:16.200 --> 00:15:18.700 Yeah, I had to learn how to use Linux on a terminal, right? 00:15:19.640 --> 00:15:19.940 Exactly. 00:15:20.060 --> 00:15:22.060 It was like, where is the UI? 00:15:22.620 --> 00:15:25.060 This is very different. 00:15:25.380 --> 00:15:27.700 How am I supposed to accomplish anything with just the terminal? 00:15:27.920 --> 00:15:31.320 And, you know, you get used to it and it's amazing in its own special way, 00:15:31.440 --> 00:15:34.280 but it doesn't feel amazing at the beginning most of the time, I think. 00:15:34.420 --> 00:15:35.520 Well, I'm reminded of the... 00:15:35.520 --> 00:15:36.140 It feels intimidating. 00:15:36.480 --> 00:15:36.600 Yeah. 00:15:36.840 --> 00:15:39.680 Every time I teach somebody how to use the Linux or the terminal, 00:15:40.440 --> 00:15:43.900 it's a big jump, I think, for people who are not familiar with CLI, 00:15:44.100 --> 00:15:46.580 which is, I think, the vast majority of people in the population. 00:15:47.320 --> 00:15:47.480 Yeah. 00:15:47.860 --> 00:15:51.919 Well, and things like people being primarily on phones for their computing device 00:15:51.940 --> 00:15:57.000 don't make that easier. It's only harder, right? This portion of Talk Python To Me is brought to 00:15:57.000 --> 00:16:01.520 you by Sentry. You know Sentry for their great error monitoring, but let's talk about logs. 00:16:02.180 --> 00:16:07.060 Logs are messy. Trying to grep through them and line them up with traces and dashboards just to 00:16:07.200 --> 00:16:13.340 understand one issue isn't easy. Did you know that Sentry has logs too? And your logs just became 00:16:13.640 --> 00:16:19.279 way more usable. Sentry's logs are trace connected and structured, so you can follow the request flow 00:16:19.300 --> 00:16:20.620 and filter by what matters. 00:16:21.320 --> 00:16:23.100 And because Sentry surfaces the context 00:16:23.400 --> 00:16:24.120 right where you're debugging, 00:16:24.380 --> 00:16:26.180 the trace, relevant logs, the error, 00:16:26.660 --> 00:16:29.380 and even the session replay all land in one timeline. 00:16:29.900 --> 00:16:31.980 No timestamp matching, no tool hopping. 00:16:32.580 --> 00:16:33.860 From front end to mobile to backend, 00:16:34.080 --> 00:16:34.700 whatever you're debugging, 00:16:34.980 --> 00:16:36.480 Sentry gives you the context you need 00:16:36.620 --> 00:16:38.740 so you can fix the problem and move on. 00:16:39.220 --> 00:16:41.520 More than 4.5 million developers use Sentry, 00:16:41.840 --> 00:16:43.720 including teams at Anthropic and Disney+. 00:16:44.420 --> 00:16:46.479 Get started with Sentry logs and error monitoring 00:16:46.500 --> 00:16:53.220 today at talkpython.fm/sentry. Be sure to use our code talkpython26. The link is in your 00:16:53.350 --> 00:16:59.720 podcast player show notes. Thank you to Sentry for supporting the show. Let's talk about your project 00:17:00.300 --> 00:17:08.199 Boundary 120. So when you started, you built this bespoke Django app for, was this for the music 00:17:08.839 --> 00:17:14.539 study or which study was this for? This was for the music IBD study. And the first version was 00:17:14.560 --> 00:17:19.640 really to track samples and to come up with an efficient way for our teams to handle the, you 00:17:19.640 --> 00:17:24.600 know, the 30,000 that we are projecting to sort out. So one of the key functionalities that we 00:17:25.060 --> 00:17:29.760 wrote in at the start was this ability to, you know, stick a QR code, stick it onto a sample 00:17:29.920 --> 00:17:34.760 label, scan it in, and that will register it into a database. And then we could move the sample 00:17:34.980 --> 00:17:39.299 around between sites and scan it and update. So that's where it started out as a sample tracking 00:17:39.320 --> 00:17:45.980 platform. So nothing too amazing. It was just

kind of a really basic operational problem that just 00:17:46.160 --> 00:17:51.820 needed to be solved. But over time, when you get samples, the next thing that happens to the sample 00:17:51.900 --> 00:17:56.660 is that it gets processed into an experimental pipeline and you get data out of it. So the 00:17:56.860 --> 00:18:02.100 lifecycle of an experimental sample then leads you on to dealing with the data that comes back. 00:18:02.400 --> 00:18:06.239 Now, because we were taking so many different types of samples, one of the key challenges that 00:18:06.260 --> 00:18:11.660 faced our team was how do you handle the data that was coming back at us? Because we have stool 00:18:11.800 --> 00:18:18.340 samples that becomes, you know, microbiome data. We have blood samples that become genomics data. 00:18:18.650 --> 00:18:22.600 And every data type that comes back just looks a little bit different, comes back in different 00:18:22.980 --> 00:18:28.520 formats, you know. And one of the difficulties as well is the size of the data that we get. 00:18:28.710 --> 00:18:34.860 So for my personal research project during that time, I was looking at sequencing out the cell-free 00:18:34.900 --> 00:18:40.160 DNA. So the DNA fragments that you have floating around your blood and each sequencing file for a 00:18:40.400 --> 00:18:44.260 participant will be about five to 10 gigabytes of data. And it comes back in a, you know, 00:18:44.560 --> 00:18:49.580 in a fast QGZ format, and that has to go into a bioinformatic pipeline. So it's the one of the 00:18:49.660 --> 00:18:54.480 key challenges is how we build that data. So Foundry 120 actually became a project that 00:18:54.720 --> 00:19:01.159 started from sample tracking into data management and handling. And in 20, you know, in the last year 00:19:01.180 --> 00:19:07.700 too, it's also become a platform where we can apply agentic AI into it to kind of accelerate 00:19:07.830 --> 00:19:12.900 our workflows. Yeah, it's a really neat platform. And I'm definitely going to dive into it. I think 00:19:13.000 --> 00:19:18.400 it's neat how you had this smaller bespoke thing and you're like, all right, let's sort of pull 00:19:18.600 --> 00:19:24.659 out the essence of it and make it useful for all kinds of scientific research. Going back to the 00:19:24.680 --> 00:19:29.360 samples. You said each one is five to six gigabytes. That's each of the 30,000? 00:19:30.340 --> 00:19:37.620 Not all the 30,000. So a subset of them, but it depends on what essays were being run on which 00:19:37.800 --> 00:19:43.140 samples. And there could be a ton of essays all running different types of samples. And each of 00:19:43.140 --> 00:19:51.279 them will have their own data type that comes back. So for my cell-free DNA work, we ran it on a subset 00:19:51.280 --> 00:19:53.260 of the population, not the entire population. 00:19:53.730 --> 00:19:57.480 But for some of the other larger scale things like genotyping, we would genotype the

whole 00:19:57.700 --> 00:19:57.940 population. 00:19:58.200 --> 00:20:04.680 So there are lots of moving parts and tiny little details here that cause so much operational 00:20:04.940 --> 00:20:05.120 problems. 00:20:06.000 --> 00:20:06.820 That's a lot of data. 00:20:07.370 --> 00:20:11.660 Where did you, how much did you end up with at the end and how did you store it and manage 00:20:11.780 --> 00:20:11.880 it? 00:20:12.080 --> 00:20:19.899 So currently we have about 10 terabytes of data and we've only, you know, processed only 00:20:19.920 --> 00:20:26.100 a fraction of those 30,000 is still being processed because some pipelines takes a human 00:20:26.740 --> 00:20:32.340 12 hours to process a couple of samples. So there's still a lot of work to be done. 00:20:32.960 --> 00:20:38.160 But yeah, it's about 10 terabytes of data that we're handling. And initially, it would be scattered 00:20:38.300 --> 00:20:42.800 across our group. Somebody's doing a microbiome project, it would be on their laptop. Otherwise, 00:20:42.940 --> 00:20:48.759 it'd be on a shared drive, on a file share somewhere in the university server. So data 00:20:48.780 --> 00:20:53.020 tends to be scattered. So if somebody is doing a specific experiment, they might create an Excel 00:20:53.240 --> 00:20:58.420 file with the readings from certain essays, and that would be the data set. And that's, I think, 00:20:58.420 --> 00:21:03.140 one of the big problems that we face in biomedical research today. Yeah, that's a lot of data. Are 00:21:03.260 --> 00:21:09.440 using things like blob storage? I know this Foundry 120 project is pretty strongly based in 00:21:09.740 --> 00:21:15.519 the Azure cloud. Are you using Azure blob storage and stuff like that? Or is it really just all on 00:21:15.540 --> 00:21:18.440 on-premise with regard to the university? 00:21:19.160 --> 00:21:22.540 So back in the days, that was the situation that we were in. 00:21:22.840 --> 00:21:25.920 But today, we're now migrated everything into Azure Blob Storage. 00:21:26.570 --> 00:21:29.220 That just gives us a scalable way of not worrying, you know, 00:21:29.730 --> 00:21:33.160 how much space we had in our shared drives. 00:21:33.920 --> 00:21:35.320 Because back in the days, you know, 00:21:35.590 --> 00:21:37.580 the university would give you a quota on your shared drive 00:21:37.670 --> 00:21:40.240 and you have to email somebody in IT to get it increased. 00:21:40.700 --> 00:21:45.080 So you might have a terabyte cap on your file share. 00:21:45.400 --> 00:21:52.880 But with Azure Blob Storage, what you end up having is just a bill at the end of the month rather than hard limits. 00:21:52.880 --> 00:21:54.260 You have a bill instead of a limit. 00:21:54.580 --> 00:21:55.080 Yeah, yeah, yeah. 00:21:55.980 --> 00:21:56.180 Interesting. 00:21:57.220 --> 00:22:03.780 When I was working way back when I was in college, in university, working in this math research lab. 00:22:03.840 --> 00:22:06.820 I told

this story a couple of times, but it's been a while, so maybe I'll share it again. 00:22:06.900 --> 00:22:13.780 And the whole of the math research department got access to the Silicon Graphics mainframe 00:22:14.340 --> 00:22:15.080 beast of a computer. 00:22:15.520 --> 00:22:19.440 We all had shared sort of workstation access to it. 00:22:19.720 --> 00:22:24.060 And one of the students, grad students, was having a problem with their code. 00:22:24.130 --> 00:22:26.120 And so they started logging out what it was doing. 00:22:26.490 --> 00:22:28.500 And they got it into an infinite loop and ran it. 00:22:28.660 --> 00:22:30.800 You would run stuff overnight and come see it in the morning. 00:22:31.140 --> 00:22:34.040 We came back one day and it just wouldn't turn on or it wouldn't respond. 00:22:34.290 --> 00:22:36.000 And nobody could figure out why. 00:22:36.380 --> 00:22:40.460 The students, we had no, the reason I'm telling you this, we had no quotas, no limits. 00:22:41.100 --> 00:22:46.240 The student used up the entire hard drive of the Silicon Graphics machine to the very last 00:22:46.440 --> 00:22:50.140 byte, and apparently it needed a few temp files to operate the operating system. 00:22:50.480 --> 00:22:51.180 And it just died. 00:22:51.560 --> 00:22:56.680 And nobody could get it to come, it took a day or two for it to come back because somebody 00:22:56.900 --> 00:22:57.300 destroyed it. 00:22:57.400 --> 00:22:59.940 Well, these limits, they have a reason they're there, you know? 00:23:00.500 --> 00:23:01.100 They do. 00:23:01.600 --> 00:23:08.140 But when we look at the next five to 10 years of biomedical research, file storage is actually a big problem. 00:23:09.040 --> 00:23:09.900 Is it? Okay. 00:23:10.100 --> 00:23:14.620 Some of the newer technologies are generating like a terabyte of data for a single sample. 00:23:14.960 --> 00:23:16.760 Wow. Yeah, that's absolutely crazy. 00:23:17.340 --> 00:23:23.320 You know, this also goes to reproducibility and long-term viability of this research, right? 00:23:23.440 --> 00:23:31.900 Because if you have 100 terabytes of data, it's one thing to say, well, here's the Excel file and here's the Parquet file and here's the Docker image that runs it. 00:23:32.300 --> 00:23:37.660 Save that as a, you know, you can stamp these as like, here's what the paper was published on, right? 00:23:37.690 --> 00:23:38.740 And here's the digital assets. 00:23:39.130 --> 00:23:42.460 But when it's 100 terabytes, it doesn't matter if you can name it or not. 00:23:42.560 --> 00:23:43.600 That's a hard thing to store. 00:23:43.920 --> 00:23:45.620 Yes, definitely a big problem. 00:23:46.700 --> 00:23:50.900 And I think the question then becomes what is the most valuable data that you store? 00:23:51.220 --> 00:23:54.280 because perhaps you don't need to store the raw files 00:23:54.540 --> 00:23:56.220 and you want to process files 00:23:56.570 --> 00:23:59.340 because who wants to run another 100 terabyte

pipeline anyway? 00:24:00.020 --> 00:24:01.600 But yeah, these are challenging questions. 00:24:01.900 --> 00:24:02.660 There are no easy answers. 00:24:03.800 --> 00:24:07.440 And most groups, we do cost our data storage, 00:24:08.080 --> 00:24:10.140 but for most grants, we cost it out for 10 years. 00:24:10.940 --> 00:24:14.220 And beyond that, it becomes a challenge to maintain all this data. 00:24:14.520 --> 00:24:15.300 Yeah. What are you going to do? 00:24:15.600 --> 00:24:18.059 I guess the only bonus is in general 00:24:18.080 --> 00:24:21.060 that storage is getting cheaper by a lot. 00:24:21.880 --> 00:24:23.700 And I say by in general, 00:24:23.810 --> 00:24:24.860 because the last couple of years 00:24:25.040 --> 00:24:26.700 or last year and a half, that's not true. 00:24:26.840 --> 00:24:28.460 Yeah, I was going to say the AI revolution 00:24:28.650 --> 00:24:32.620 might not keep that falling price curve the same. 00:24:33.520 --> 00:24:35.780 It might change things just a little bit. 00:24:36.160 --> 00:24:38.120 All right, so let's talk about this Foundry project. 00:24:38.210 --> 00:24:40.000 So this is sort of the next generation 00:24:40.240 --> 00:24:42.700 of what you maybe dreamed of building. 00:24:44.380 --> 00:24:47.520 Also maybe a little bit in the agentic age as well, right? 00:24:47.580 --> 00:24:48.080 Yeah, definitely. 00:24:48.740 --> 00:24:50.700 Okay, so tell us about this Foundry 120. 00:24:51.100 --> 00:24:51.820 So Foundry 120. 00:24:51.900 --> 00:24:53.580 People will find it at foundry120.com. 00:24:54.140 --> 00:24:57.400 It's a research operating system for translational science teams. 00:24:57.580 --> 00:24:58.980 So I've got to explain that a bit. 00:24:59.560 --> 00:25:04.540 So translational science is basically when you're trying to discover things by recruiting 00:25:05.150 --> 00:25:09.820 humans with a problem and you're taking samples, bringing them back to the lab and running all 00:25:09.960 --> 00:25:14.780 sorts of experiments to just kind of understand the biology behind disease and then find out 00:25:14.780 --> 00:25:18.820 whether there are mechanisms that we can target for therapeutic development and things like that. 00:25:19.480 --> 00:25:20.920 So that's kind of translational science. 00:25:21.190 --> 00:25:26.160 And Foundry 120 is a platform that allows teams to operate seamlessly 00:25:26.790 --> 00:25:29.200 and use AI to accelerate their workflows. 00:25:29.800 --> 00:25:31.060 It's built around three ideas. 00:25:31.600 --> 00:25:33.400 The first idea is to organize everything. 00:25:33.610 --> 00:25:38.980 So you organize your samples, your studies, your participants into a database 00:25:39.380 --> 00:25:42.020 that just helps you keep track of everything. 00:25:42.440 --> 00:25:46.980 And then the second bit is the idea of centralizing all your data outputs into a single platform. 00:25:47.440 --> 00:25:53.640 So this will be things like clinical data, including radiology images, endoscopy videos, 00:25:54.360 --> 00:25:59.780 digital pathology slides, and then the scientific outputs such as spatial transcriptomics, 00:26:00.240 --> 00:26:05.620 genomics, microbiome, or all the data volume that we're seeing from all the different scientific 00:26:06.000 --> 00:26:11.080 modalities are just going up exponentially. But as we know in the AIH, if you can bring all your 00:26:11.080 --> 00:26:18.480 context into a single platform, then you can let AI connect to that and operate on that to answer 00:26:18.720 --> 00:26:24.520 questions that researchers might have. So, I mean, I'll give you a concrete example. For example, 00:26:24.730 --> 00:26:29.620 you know, if one of my scientific colleagues were looking for, you know, do we have any samples that 00:26:29.720 --> 00:26:35.420 belong to a participant who's been treated with this drug? Can we find that? And historically, 00:26:35.480 --> 00:26:41.220 what you would have to do is to take the clinical data set, join it with your sample database, 00:26:41.900 --> 00:26:47.020 and then filter that through and find the samples that you want. And that usually would have taken 00:26:47.060 --> 00:26:51.980 a couple of days to do. But this is a great use case, I think, for agentic AI, where you can just, 00:26:52.600 --> 00:26:58.140 you know, access the data, put it into a sandbox, write the code to do the join, and then just give 00:26:58.140 --> 00:27:03.659 you the answer that you want. So just eliminate all the steps in between and get you to the answer 00:27:03.680 --> 00:27:08.700 faster. And that's the whole concept behind the platform. Sounds great. It's a really nice looking 00:27:09.180 --> 00:27:15.060 web app too. What's its current status? It's available not just for you all, but for others, 00:27:15.940 --> 00:27:22.080 but it doesn't have quite a just create an account, get started. So you have a request a platform 00:27:22.420 --> 00:27:28.380 walkthrough. What's the situation here? Can other research teams use it? Is it a paid product? Is it 00:27:28.280 --> 00:27:34.420 just sort of gated because it's not ready for people overwhelming it? So it's quite early stage 00:27:34.440 --> 00:27:40.720 at the moment. So over the last six to 12 months, we've made the entire foundation generalizable 00:27:41.280 --> 00:27:46.560 beyond just our group and beyond just disease. So what I mean by that is, you know, for another 00:27:46.620 --> 00:27:51.280 group to join, they would need the access controls to be in place. So they need a way of managing 00:27:51.720 --> 00:27:58.240 their team, being able to assign permissions to individuals, you know, without me having to do it 00:27:58.480 --> 00:28:13.320 But I think more importantly is that onboarding a team into this platform is actually quite a hands-on process because it really depends on what data you're handling, the volumes that you're dealing with, what modalities of data you need to do. 00:28:13.660 -->

00:28:20.880 And also the other major thing that we have to look at if you want to join the platform is the governance around your data. 00:28:21.420 --> 00:28:28.160 So in clinical research, there are strict rules around how you handle data and where that data can sit, where can it be processed. 00:28:28.840 --> 00:28:33.100 And these are subject to your research ethics approval. 00:28:33.810 --> 00:28:40.280 So before a team can join us, we do want to review all those things to make sure that we're in compliance before anybody can join. 00:28:40.500 --> 00:28:46.940 So that's the reason why it's not just a simple, you know, I can sign up, pay a monthly fee and start using the platform. 00:28:47.440 --> 00:28:52.120 There's a lot more issues that have to be looked at before we can bring a team on board. 00:28:52.320 --> 00:28:58.860 But we're willing to do the work to kind of review the situation and start getting other people access to it. 00:28:59.000 --> 00:28:59.720 Makes a lot of sense. 00:28:59.910 --> 00:29:02.460 You've got IRB research stuff. 00:29:02.820 --> 00:29:06.500 You've got HIPAA and the equivalent in all the other countries. 00:29:07.200 --> 00:29:09.800 And yeah, so you don't want to get in trouble. 00:29:10.000 --> 00:29:10.260 Yes. 00:29:10.570 --> 00:29:14.360 And you did mention earlier that this is quite tied into Azure at the moment. 00:29:14.900 --> 00:29:18.960 And that's because we run on the University of Glasgow's Azure Tenancy. 00:29:19.330 --> 00:29:25.400 So everything has to be located specifically for our studies within the UK data centers and things. 00:29:25.540 --> 00:29:27.780 So we have all the controls in place to do that. 00:29:28.330 --> 00:29:32.600 And that's backend infrastructure for the people who are listening on the podcast. 00:29:34.520 --> 00:29:38.000 This portion of Talk Python is brought to you by Talk Python Courses. 00:29:38.380 --> 00:29:39.780 Here's the thing that always bug me. 00:29:40.040 --> 00:29:44.620 You finish one of our courses, that's hours of video, a pile of code you actually wrote, 00:29:44.740 --> 00:29:48.140 and real skills you didn't have a month before, and then nothing happens. 00:29:48.480 --> 00:29:51.120 No paper, no credential, nothing to show for it. 00:29:51.660 --> 00:29:52.300 So we fixed it. 00:29:52.740 --> 00:29:56.440 Every Talk Python course now generates a completion certificate automatically. 00:29:56.880 --> 00:29:58.840 Go to your account page in your dashboard section, 00:29:59.380 --> 00:30:02.500 scroll down to your completed courses, and click Certificate. 00:30:02.820 --> 00:30:03.620 That's the whole process. 00:30:04.600 --> 00:30:07.200 Two things you can do with these course completion certificates. 00:30:07.560 --> 00:30:12.640 download the full PDF, which is handy if your employer reimburses training or gives you credit 00:30:12.650 --> 00:30:18.300 for finishing it. Or you can make the certificate public and hit share on LinkedIn, which

adds it to 00:30:18.300 --> 00:30:22.700 your LinkedIn profile under licenses and certifications. Not a poster that scrolls away 00:30:22.710 --> 00:30:27.360 in a day, an actual credential sitting on your profile where your manager and recruiters can see 00:30:27.440 --> 00:30:31.980 it. Plus, if you've been taking our courses for a while, you've probably earned several of these 00:30:32.090 --> 00:30:37.440 without even knowing they existed. Just visit training.talkpython.fm/account and collect 00:30:37.460 --> 00:30:42.640 them. Thanks to all of you who have taken a Talk Python course. It's a great way to support the podcast. 00:30:44.280 --> 00:30:49.260 What a different time it is for universities. And I think they're going through a similar 00:30:49.980 --> 00:30:54.400 upheaval, I guess is the word. When I was in university, and we're talking like 90s, 00:30:54.840 --> 00:31:00.680 there was a giant Cray supercomputer or maybe a couple of supercomputers at the heart of the 00:31:00.940 --> 00:31:05.019 university in some basement, and you could get access to that if you needed mega computing 00:31:05.020 --> 00:31:12.120 resources. And now it's just scaled into some ginormous cloud infrastructure. And the limit 00:31:12.130 --> 00:31:16.440 is really just how much are they going to give you, not what does the university have in its 00:31:16.620 --> 00:31:20.940 basement or something like that, which is really interesting. Well, I mean, there are still some 00:31:21.220 --> 00:31:25.700 universities that are trying to build their own clusters. But I think if you look at the economics 00:31:25.920 --> 00:31:32.659 of it, it is way more cost efficient to be running on the cloud compared to on the cluster, because 00:31:32.560 --> 00:31:37.760 many of the analysis that we do in the universities once off. So you might process a ton of data, 00:31:38.130 --> 00:31:42.360 but that might be after a year or two years before you get that amount of data to process. 00:31:42.880 --> 00:31:46.560 So if you look at a workload, the IT workload within the university, I think the cloud is 00:31:46.570 --> 00:31:50.560 quite an attractive proposition. That's an interesting angle. Yeah, of course, because 00:31:51.100 --> 00:31:55.160 you're going to process your samples. Maybe you're like, ah, we really want to try a different 00:31:55.500 --> 00:31:59.860 algorithm and you'll run it all through again. But generally speaking, it's not steady state. 00:32:00.360 --> 00:32:11.500 Whereas if you're going to go buy your own hardware and then put it in a basement and it's kind of a steady state situation, then you could really predict, you know, we're saving, you know, 50% less expensive to run on this. 00:32:11.570 --> 00:32:13.280 And just, yeah, didn't really think about that. 00:32:13.280 --> 00:32:17.880 And so I said, there's another such wave coming because this was, you know, when was the cloud? 00:32:17.950 --> 00:32:20.980 The cloud is 10 years ago, 2008, I think. 00:32:21.140 --> 00:32:24.620 A little bit, a little bit, really caught going a little bit after that. 00:32:25.150 --> 00:32:27.320 But it's been a little while. 00:32:27.360 --> 00:32:45.740 Now we've got all this AI stuff going and we're proving, you know, unsolved airdosh mathematical problems and all sorts of things, you know, with AI and other, you know, trying to do protein folding and other things that are really computational, but also really different. 00:32:45.970 --> 00:32:54.340 And so I feel like this whole AI wave, especially the agentic AI, is going to roil what's happening at the universities all over again. 00:32:54.480 --> 00:32:58.780 I mean, if you look at the data center economics around AI, it's quite interesting, actually. 00:32:59.150 --> 00:33:02.400 Like a single rack from Nvidia is like a couple million dollars, right? 00:33:03.240 --> 00:33:03.980 It's crazy. 00:33:04.130 --> 00:33:06.640 And you need a power plant to plug it into. 00:33:07.500 --> 00:33:11.400 I can't see any university building their own AI data center in the future. 00:33:11.590 --> 00:33:17.900 So I think cloud adoption is inevitable in the sense that the economics of trying to 00:33:17.900 --> 00:33:23.400 use agentic AI doesn't make sense to try and build your own AI computer, unless you have 00:33:23.360 --> 00:33:30.220 specific governance requirements, which do apply to certain studies and certain aspects of work. 00:33:30.940 --> 00:33:35.980 But even then, you might contract a local AI company to do it and then share it between 00:33:36.400 --> 00:33:40.400 businesses and universities rather than having a university do it on its own. 00:33:40.760 --> 00:33:45.520 Yeah. People talk about the AI bubble. I don't know if there's actually an AI bubble. This stuff 00:33:45.660 --> 00:33:51.559 is so productive. The last bubble we had that was tech-related was the dot-com bubble. And there 00:33:51.480 --> 00:33:57.160 was really stupid stuff with a lot of money spent on it you know like there's always the pest.com or 00:33:57.170 --> 00:34:01.900 the weird investments you know thing and people are spending millions of dollars on ads to just 00:34:02.120 --> 00:34:06.520 like have dancing monkeys running it was a weird time and this is also a weird time but I feel like 00:34:06.840 --> 00:34:12.120 there's actually something legitimately at the core of it that is really changing the way people 00:34:12.300 --> 00:34:17.300 work and and do research and so on so I don't know it's going to bust if there is a bubble and I don't 00:34:17.240 --> 00:34:22.240 know if the bubble is going to burst. But if it does, I think that's also going to create an 00:34:22.540 --> 00:34:27.879 explosion of local AI. You know, all of a sudden, right now, if you have to pay $100 for your 00:34:28.260 --> 00:34:32.520 Anthropic subscription, that's totally reasonable. But if that becomes a $2,000 a month bill, 00:34:32.879 --> 00:34:37.940 well, then

buying a $10,000 workstation that can do legit local AI is all of a sudden a bargain, 00:34:38.139 --> 00:34:43.139 you know what I mean? So it could be in the future that there's sort of a coming back to 00:34:43.159 --> 00:34:47.899 local compute as well for AI, maybe. Yeah, I mean, the small models are getting more and more capable 00:34:48.159 --> 00:34:52.980 as well. So not every workload needs a frontier model at the moment. So what you're saying is 00:34:53.080 --> 00:34:56.919 probably true in the sense that there is definitely going to be a role for local AI. 00:34:57.230 --> 00:35:02.560 I think Gemma maybe works this way, but certainly I can also see a world where there's kind of an 00:35:02.780 --> 00:35:10.099 orchestration layer and then 100 or 50 specialists that get selected, right? Right now, if you ask a 00:35:10.060 --> 00:35:15.940 frontier models to do something on genomics, it uses the same giant model as it would to use to 00:35:16.280 --> 00:35:22.840 like write Shakespeare derived things. Right. But if you had one that just is trained on genomics, 00:35:23.920 --> 00:35:27.380 you could have a much smaller model that could run locally and you could say, okay, 00:35:27.660 --> 00:35:31.400 this part of the question goes to this, this sub model. And I don't know, it's going to be 00:35:31.540 --> 00:35:35.420 interesting where it goes. And that let's, so coming back to it, let's, let's talk a little 00:35:35.400 --> 00:35:40.660 bit about Foundry. There's a little demo you've got going right here at the beginning. Let me see 00:35:40.660 --> 00:35:46.640 if I can go to the front of it. And maybe, I guess there's the stuff you talked about before, 00:35:46.780 --> 00:35:53.460 the sample gathering and organizing and all that kind of stuff, right? And then on the back of that, 00:35:53.700 --> 00:36:00.260 once you get it all done, there's this local tool using AI that understands all of your research 00:36:00.280 --> 00:36:06.920 data. So maybe tell us about the data collection and the non-AI bit, and then it'll be fun to talk 00:36:07.020 --> 00:36:11.240 about the AI because it's kind of non-standard. It's kind of powerful. Yeah, let's talk about the 00:36:11.320 --> 00:36:15.260 data bit because that has been one of the most difficult problems I've been thinking about for 00:36:15.300 --> 00:36:22.160 the last few years is how do we aggregate all these types of data sets into some kind of a common 00:36:22.840 --> 00:36:29.640 model that an AI... Well, nowadays we can use AI, but it still applies to humans. So when my 00:36:29.660 --> 00:36:36.420 colleagues are looking for data and trying to match, you know, data set A to B, what is the 00:36:36.580 --> 00:36:42.000 way we do it? And at the end of the day, I think, you know, we've brought down everything to a model 00:36:42.200 --> 00:36:48.400 where all data exists in files. So even tabular data, we keep it in CSV file. Well, historically, 00:36:48.840 --> 00:36:52.880

we've tried to put, you know, tabular data into the database. But I think going forward, 00:36:53.200 --> 00:36:58.080 we're just going to keep the files because that's the universal denominator. So whether you're 00:36:58.100 --> 00:37:04.960 dealing with endoscopy videos, MRI images, digital pathology slides. Everything's a file, and each 00:37:05.050 --> 00:37:11.020 file has each set of, you know, each type of data has its own file structure, but the common language 00:37:11.310 --> 00:37:16.700 at the end of day is files. So we put files, and the thing that makes the files useful is putting a 00:37:16.810 --> 00:37:21.720 layer of metadata and relationship data on top of the file. So I think by coupling the two of them, 00:37:21.960 --> 00:37:25.920 we found this common model that can generalize across all the data sets that we use. 00:37:26.120 --> 00:37:46.640 Very interesting. And, you know, to your point of just keep it in the files, you know, with things like DuckDB and other really cool things and Parquet files, you can kind of treat them like databases already. So I don't know if people know, but with DuckDB, you can say things like select star from read Parquet input, you know, which is pretty insane. 00:37:47.200 --> 00:37:49.120 Yeah, Parquet is great. We use it as well. 00:37:49.840 --> 00:38:00.740 Yeah. So also things like, what do you think about little SQLite files or DuckDB file where there's these just embedded no server databases? 00:38:01.440 --> 00:38:04.840 And I think there's probably some really good use cases and research for that. 00:38:05.100 --> 00:38:10.700 Potentially, you know, SQLite's a very interesting concept of having an entire database in a file. 00:38:13.020 --> 00:38:18.240 But when I think of my end users, my colleagues who are running science experiments and such, 00:38:18.760 --> 00:38:21.100 they're used to dealing with Excel and CSV files. 00:38:21.690 --> 00:38:25.680 So we try to keep the same file format everybody's familiar with. 00:38:25.920 --> 00:38:30.140 But I think for some of the future work that we're going to do, we can use some of these 00:38:30.770 --> 00:38:32.060 more specific file formats. 00:38:32.690 --> 00:38:37.560 Because as you know, if everything's a file, then an AI agent can run across all the files 00:38:37.820 --> 00:38:40.480 and pull out the data that it needs for what it needs to do. 00:38:40.700 --> 00:38:41.240 Yeah, absolutely. 00:38:41.490 --> 00:38:41.840 Yeah, sure. 00:38:41.940 --> 00:38:48.920 If you're giving them the files directly, here's either your Excel workbook or here's your CSV file. 00:38:49.110 --> 00:38:56.980 But if it's coming in and out of a platform like Foundry 120, you can store it as one thing and then export it or import it as another, right? 00:38:57.140 --> 00:39:03.620 Yeah. I mean, so fundamentally, what we do is that the files sit in Azure Blob Storage. 00:39:04.360 --> 00:39:11.120 And when Helix, our AI agent, wants to process a file, it spins up a VM. 00:39:11.660 --> 00:39:14.560 The file gets transferred from the Azure storage into the VM. 00:39:15.000 --> 00:39:20.680 The AI sends its analysis code, whether it's in Python or any other language, into the VM. 00:39:21.010 --> 00:39:27.400 The computation happens, and then we return the output to Helix and store the output back in Azure storage. 00:39:28.200 --> 00:39:34.820 So by doing that, we put the sandbox guardrail around the AI agent for security purposes. 00:39:35.560 --> 00:39:39.460 but it also allows us to enable the AI to do processing, 00:39:40.060 --> 00:39:41.580 generate new files, store it. 00:39:41.860 --> 00:39:45.000 And I think that's the kind of model that works for us in Foundry 00:39:45.640 --> 00:39:48.060 while keeping it all within an institutional Azure environment 00:39:48.340 --> 00:39:51.760 and also allows us to enforce all the role-based access controls 00:39:51.820 --> 00:39:53.540 that we need for our team members. 00:39:54.060 --> 00:39:57.760 So when they query the AI, the AI can only see the same files 00:39:57.940 --> 00:40:00.460 that they would have normal access to. 00:40:01.040 --> 00:40:04.700 And I think that's the design that's very specific for this 00:40:04.720 --> 00:40:07.980 because of the governance requirements that we have around the data sets that we use. 00:40:08.400 --> 00:40:09.440 These AIs are sneaky. 00:40:10.180 --> 00:40:12.680 They will find a way to access the files. 00:40:12.880 --> 00:40:16.260 I mean, the really big headline cases are like, 00:40:16.660 --> 00:40:19.760 OpenAI was training its model and it hacked multiple systems 00:40:20.060 --> 00:40:22.520 so that it could get to the answers on Hugging Face 00:40:22.880 --> 00:40:25.840 instead of just figuring out the, solving the test, right? 00:40:25.860 --> 00:40:29.680 It's like a teenager that doesn't really care about the work is just doing it. 00:40:29.860 --> 00:40:33.079 But what I was thinking was, you know, I was working with Claude 00:40:33.100 --> 00:40:35.820 and I had some question about my code 00:40:35.850 --> 00:40:37.640 and it said something to the effect of like, 00:40:37.690 --> 00:40:39.600 oh yeah, you have two GitHub issues on this. 00:40:40.020 --> 00:40:42.060 I never gave it direct access to GitHub. 00:40:42.100 --> 00:40:43.160 I never gave it access to GitHub. 00:40:43.230 --> 00:40:44.360 I'm like, how does it know that? 00:40:45.160 --> 00:40:47.180 It's quoting like GitHub stuff, 00:40:47.400 --> 00:40:49.360 not through Git history locally, 00:40:49.540 --> 00:40:51.520 but it like reading the issues and the PRs. 00:40:51.520 --> 00:40:52.280 I'm like, how does it do? 00:40:52.340 --> 00:40:54.740 And then I realized I had the GitHub CLI installed 00:40:54.940 --> 00:40:55.600 and it's like, well, 00:40:56.000 --> 00:40:57.480 let me see if the GitHub CLI is installed. 00:40:57.610 --> 00:41:00.220 Oh, and look, it's already automatically authenticated 00:41:00.360

--> 00:41:02.540 because Michael logged in at some point 00:41:02.560 --> 00:41:07.420 to the CLI. And so it was just using the CLI that it's also, you know, like, oh, let me check that 00:41:07.480 --> 00:41:11.120 on the server for you. I'm like, excuse me. Yeah. Yeah. On your production server, you're doing this. 00:41:11.200 --> 00:41:15.440 How you're not supposed to be there. Why? And you know, it's just like, it's realized in the code 00:41:15.560 --> 00:41:20.920 somehow it's figured out that it can SSH. So you got to be really careful about those things. Right. 00:41:21.120 --> 00:41:25.580 Cause they're not malicious. They're just like, you asked me to solve a problem. And if I can get 00:41:25.610 --> 00:41:30.540 to that, I got a better answer, more concrete data. And, but it could also go and we fixed the problem 00:41:30.560 --> 00:41:31.760 by resetting the database. 00:41:31.990 --> 00:41:33.340 Like, oh, no, you didn't. 00:41:34.980 --> 00:41:36.380 Yeah, so for us... 00:41:36.440 --> 00:41:39.600 So how do you do that kind of stuff in your project? 00:41:39.820 --> 00:41:40.680 So it's the backend. 00:41:40.830 --> 00:41:42.660 So Django enforces the permissions. 00:41:44.100 --> 00:41:49.260 And actually, so when the AI stages data and stages code, 00:41:49.410 --> 00:41:51.080 that actually goes through Django first. 00:41:51.660 --> 00:41:54.500 So the AI is not calling directly into the files, 00:41:54.500 --> 00:41:56.440 is not calling directly into a compute environment. 00:41:56.940 --> 00:42:00.520 And then Django screens all that code 00:42:00.540 --> 00:42:02.440 and activates the VM. 00:42:02.780 --> 00:42:05.820 So we've actually put Django in as a security guard 00:42:06.300 --> 00:42:10.960 between your AI and the raw data and the compute environments 00:42:11.060 --> 00:42:11.600 that we run in. 00:42:11.820 --> 00:42:12.580 MARK MANDEL: Oh, very cool. 00:42:12.760 --> 00:42:15.720 So Foundry 120 is also written on Django, 00:42:16.080 --> 00:42:20.200 but it's more Django REST framework and TypeScript React. 00:42:20.460 --> 00:42:21.620 Is that the story? 00:42:21.620 --> 00:42:22.360 FRANCESC CAMPOY: Yep, that's right. 00:42:22.520 --> 00:42:23.760 So on the front end, it's- 00:42:23.760 --> 00:42:24.640 MARK MANDEL: Yeah, so tell us a bit about it. 00:42:24.640 --> 00:42:28.660 FRANCESC CAMPOY: So the old app that we used to run off 00:42:28.740 --> 00:42:30.140 was just pure Django. 00:42:30.360 --> 00:42:34.960 So we use Django templates to handle all the registration and the CRUD workflows. 00:42:36.100 --> 00:42:40.540 But as we move into this age of AI, AI, as you know, is pretty asynchronous. 00:42:40.970 --> 00:42:45.340 So every call you make takes, you know, sometimes it feels like forever to come back with a response. 00:42:46.420 --> 00:42:52.940 And when we think about the async nature of all the calls that have to be made to run an AI conversation or agent loop, 00:42:53.820 --> 00:42:58.720 TypeScript and JavaScript tends to come to be more suitable for that kind of application. 00:42:59.720 --> 00:43:00.720 So we run both. 00:43:01.130 --> 00:43:03.400 So we have TypeScript on the front end, Django on the back end. 00:43:03.470 --> 00:43:04.400 I think it's a great setup. 00:43:04.510 --> 00:43:09.940 It gives us access to the entire Python data science ecosystem, while also giving us all 00:43:10.080 --> 00:43:15.800 the TypeScript and JavaScript ecosystem for handling all these asynchronous work and the 00:43:16.280 --> 00:43:19.440 interactivity that we want on the front end when you start running AI applications. 00:43:20.040 --> 00:43:20.400 Makes sense. 00:43:20.760 --> 00:43:25.520 Now, what I'm about to ask you doesn't really make sense because of the AI angle and that 00:43:25.570 --> 00:43:26.040 kind of stuff. 00:43:26.240 --> 00:43:46.520 But if you think about scaling this out to other research projects and other groups, have you considered looking at things like PyOxid, Iodide, sorry, and things like JupyterLite for running some of that compute on the front end on people's browsers so that you don't have to basically pay the compute cost? 00:43:47.880 --> 00:43:56.560 So that's an interesting question because now you're asking me about the compute architecture that we have in the backend, which is actually pretty heavy. 00:43:57.790 --> 00:44:08.320 So we've mentioned that some of the files that we have might be gigabytes in size and the compute power you need to run genomics pipeline is pretty high. 00:44:08.840 --> 00:44:15.800 So for a concrete example, so in the backend of our Foundry system, we have access to about 350 CPUs on Azure. 00:44:16.120 --> 00:44:29.240 So if you ask Helix for a very heavy analysis, say on a big transcriptomic data set of something, through Django, we are able to orchestrate up a heavy compute job, which will then go off and run. 00:44:29.660 --> 00:44:36.360 It'll spin up as many CPUs as it needs to, to process the job, and then returns the output later on once it's all done. 00:44:36.580 --> 00:44:39.960 And that process can take half an hour, a couple of hours. 00:44:40.330 --> 00:44:42.000 So it might come back really late. 00:44:42.260 --> 00:44:47.740 And because of that model that we're running, actually, the kind of computational requirements we have is pretty high. 00:44:48.830 --> 00:44:53.540 And so we don't really want to run compute on people's laptops. 00:44:53.890 --> 00:44:57.640 We want to run it in the cloud to handle the data sets that we're dealing with. 00:44:57.830 --> 00:44:59.400 So that's a very specific design choice. 00:45:00.520 --> 00:45:06.080 And it's all to do with the kind of data that we're handling and the need to throw a lot of RAM and a lot of CPU at it. 00:45:06.280 --> 00:45:11.260 Sure. And if your individual files are five gigs, that's a lot of just bandwidth costs. 00:45:11.360 --> 00:45:18.320 So it's like every time you want to load something,

you got to pull that five gigs out of the cloud, which has different costs and so on. 00:45:18.340 --> 00:45:22.460 Right. Well, yeah. And many of the biological data problems are parallel. 00:45:22.780 --> 00:45:25.240 Right. So you've got 200 participants each of five gigs. 00:45:25.720 --> 00:45:28.520 You might as well spin up 200 machines and process all of them in parallel. 00:45:29.120 --> 00:45:31.980 So all these problems that we have are very paralyzable. 00:45:32.280 --> 00:45:39.040 And the cloud is a great platform to do that in because you can do it on demand, on the fly, spin it up, finish processing and tear it all down. 00:45:39.600 --> 00:45:42.380 So minimal cost for, you know, maximum impact. 00:45:42.820 --> 00:45:44.520 That's what we're going for in the back end. 00:45:44.860 --> 00:45:47.300 Right. That's the bursting component that you talked about. 00:45:47.680 --> 00:45:53.860 I do think JupyterLite is pretty interesting with the local Piodide execution and all that kind of stuff. 00:45:54.160 --> 00:45:56.300 Just the fact that that's possible is it's pretty neat. 00:45:56.300 --> 00:45:59.340 But yeah, I can see that it really doesn't apply for what you're doing here. 00:45:59.640 --> 00:46:04.980 Well, I haven't talked to you about the front end of sample operations, right? 00:46:05.540 --> 00:46:11.400 Because we're running a sample collection in the hospitals across Scotland. 00:46:11.530 --> 00:46:14.520 And I must say the frontline IT infrastructure is not always the best. 00:46:14.910 --> 00:46:21.180 So by having all our compute power on the server side, we can guarantee a speedy experience for our teams working at the front end. 00:46:21.480 --> 00:46:24.960 So we're not depending on the front end's computational power. 00:46:25.340 --> 00:46:30.660 I'll tell you what, my experience looking over the shoulder at the software that doctors and nurses use, 00:46:31.180 --> 00:46:34.640 there's a lot of room for improving the user experience. 00:46:35.820 --> 00:46:36.200 Oh, definitely. 00:46:36.700 --> 00:46:40.260 I mean, part of the reason we went Django as well is because it was server-side. 00:46:40.620 --> 00:46:45.700 And, you know, on the front end, you know, we have some computers I've used in hospitals. 00:46:46.520 --> 00:46:48.440 They go back to 2010, right? 00:46:48.740 --> 00:46:51.440 You know, we're running on Intel chips like 20 years old. 00:46:52.620 --> 00:46:53.180 Oh, my goodness. 00:46:54.000 --> 00:47:18.280 Yep. And there's a lot of Cisco, remote, whatever there. So let's talk about the AI side now. So we talked about the data, the data handling, some of the tech behind Foundry 120. But I think one of the cornerstones is this Helix AI. Now, when people think there's a bit of a problem here, Sean, like people talk about AI, and there's two or three different things it could be, and they all use the same word. 00:47:18.580 --> 00:47:22.580 And they think they're talking about the same thing, but they're actually talking past each other. 00:47:22.830 --> 00:47:27.220 You know, like I asked ChatGPT for this math problem and it got it wrong. 00:47:27.460 --> 00:47:32.240 It's like, yeah, but we also built incredible software with this other thing that we also call AI. 00:47:32.600 --> 00:47:35.980 And, you know, it's always right because it writes Python to actually answer its questions. 00:47:36.090 --> 00:47:37.980 And yeah, it's just really interesting. 00:47:38.360 --> 00:47:48.420 So there's an AI that you've mentioned a couple of times in here that will help researchers ask questions, find data, look for trends and those kinds of things. 00:47:48.700 --> 00:47:57.480 And this, I think, you give me your thoughts on this, but my feeling is that it's a little bit like a Claude code or a codex. 00:47:57.660 --> 00:48:04.460 One of these sort of tool using self-correcting AIs, not just a chat LL. 00:48:04.480 --> 00:48:11.840 Yeah. So Helix is an agentic AI system. And I think the problem that you're describing is 00:48:12.020 --> 00:48:17.200 because most people's experience of AI is chatbots. You go to chatgpt.com, you ask a question, 00:48:17.550 --> 00:48:21.880 it gives you an answer. But actually the stuff that we're seeing that makes us think that AI 00:48:22.060 --> 00:48:26.960 might not be a bubble is all this agentic AI stuff that we're seeing. So Claude Code, codecs, 00:48:27.780 --> 00:48:33.700 agentic AI is a very different paradigm from chatbots, right? In agentic AI, the AI, 00:48:33.980 --> 00:48:38.240 you give the AI a task, it looks at its tool set, it looks at what you're trying to do, 00:48:38.450 --> 00:48:42.520 and then it goes away and works at it until it gives you an answer. And that's incredibly powerful. 00:48:42.900 --> 00:48:48.040 Whereas I think, you know, 95% of people's experience of AI is almost like a Google search. 00:48:48.230 --> 00:48:52.700 You go to ChatGPT and you ask, hey, what's the directions to this place or what's the recipe for 00:48:52.820 --> 00:48:57.880 that? And therefore, there's this huge gap in understanding of how powerful agentic AI systems 00:48:57.900 --> 00:49:03.880 can be. So Helix is really one of the, we think it's one of the first demonstrations of how you 00:49:03.960 --> 00:49:10.240 would apply agentic AI in the science world. And a lot of the scientists that I'm showing this 00:49:10.480 --> 00:49:14.860 system to, this is the first time that they're seeing an agentic AI system. So, you know. 00:49:15.120 --> 00:49:15.800 What's their reaction? 00:49:18.420 --> 00:49:22.540 I think everybody's quite excited. They're like, oh, that used to take me like months to do, 00:49:22.700 --> 00:49:24.420 or it took me a lot of emails to do. 00:49:24.680 --> 00:49:25.900 Just doing it in minutes right now, 00:49:26.800 --> 00:49:28.140 it's quite amazing. 00:49:28.400 --> 00:49:30.300 And historically, every team member 00:49:30.420 --> 00:49:31.100 that's joined our

team, 00:49:31.240 --> 00:49:32.400 I've had to sit down with them 00:49:32.470 --> 00:49:34.580 and teach them how to use Python or R 00:49:34.630 --> 00:49:37.880 to get a graph out of Excel file or something. 00:49:38.280 --> 00:49:40.340 And now you can just hand it off to Helix, 00:49:40.680 --> 00:49:42.660 let it do it, let it write the code. 00:49:42.910 --> 00:49:45.020 And then what you do in this situation 00:49:45.070 --> 00:49:47.040 is that you've got to verify that it's correct. 00:49:47.520 --> 00:49:48.880 So the work changes. 00:49:49.050 --> 00:49:50.500 So instead of you as a researcher 00:49:50.800 --> 00:49:51.740 writing the analysis code, 00:49:52.040 --> 00:49:54.060 You get the AI to do the analysis for you. 00:49:54.220 --> 00:49:56.280 And then what you have to do is verify that it is correct. 00:49:56.660 --> 00:49:58.160 So it's a very different way of working. 00:49:58.860 --> 00:50:00.240 But it's incredibly powerful. 00:50:00.460 --> 00:50:01.420 It's much faster. 00:50:01.660 --> 00:50:04.680 It's taken a lot of road work out of everybody's life. 00:50:04.840 --> 00:50:06.460 So I think everybody's really excited about it. 00:50:06.680 --> 00:50:07.140 I would imagine. 00:50:07.880 --> 00:50:10.900 I'm going to have you talk us through just this sort of workflow that it goes through 00:50:10.960 --> 00:50:11.320 real quick. 00:50:11.620 --> 00:50:15.940 But that's the big danger is that it just hallucinates, which I don't know, that's a 00:50:16.040 --> 00:50:16.340 weird word. 00:50:16.400 --> 00:50:20.920 It's just it's wrong, whatever, about the actual data. 00:50:21.260 --> 00:50:25.500 Do you all use RAG, the sort of training on the data, 00:50:25.540 --> 00:50:29.460 or is it just really the tool-using components that make it go? 00:50:30.420 --> 00:50:31.980 It depends on what you mean by RAG, 00:50:32.220 --> 00:50:35.500 because we use tools to ground the AI. 00:50:36.840 --> 00:50:40.860 Yeah, I'm thinking like going and actually retraining the model 00:50:41.120 --> 00:50:43.500 on the research files and data, 00:50:43.640 --> 00:50:45.300 which I'm guessing from looking at it, 00:50:45.480 --> 00:50:46.360 it doesn't look like it. 00:50:46.360 --> 00:50:48.960 It looks more of a Claude Code tool-using style. 00:50:49.060 --> 00:50:50.740 Yeah, it is more Claude Code style. 00:50:52.480 --> 00:50:55.980 we don't train the AI specifically for it. 00:50:56.190 --> 00:50:59.280 We can swap the base models as newer versions come up. 00:50:59.980 --> 00:51:05.580 Because if supervised fine-tuning or doing some RL on a base, 00:51:06.020 --> 00:51:07.420 LLM will cost you a lot of money. 00:51:09.180 --> 00:51:10.660 And it's not generalizable, right? 00:51:11.160 --> 00:51:14.900 Yeah, and you need the data to train it on. 00:51:15.350 --> 00:51:17.200 And that's not easy to make as well. 00:51:18.620 --> 00:51:19.860 So yeah, so given the rapid progress, 00:51:20.760 --> 00:51:23.440 we want a model where

we can just update to the latest model 00:51:23.920 --> 00:51:27.840 and just leverage the latest changes 00:51:28.080 --> 00:51:29.640 that the big labs are coming out with. 00:51:30.360 --> 00:51:30.580 Amazing. 00:51:30.860 --> 00:51:31.700 I think that's the right way. 00:51:31.820 --> 00:51:34.560 So if you go to foundry120.com, 00:51:34.920 --> 00:51:36.800 there's a one-minute little screencast, 00:51:37.020 --> 00:51:38.320 silent screencast of it going. 00:51:38.600 --> 00:51:40.300 So I'm going to, Sean, I'm going to hit play 00:51:40.620 --> 00:51:42.240 and you kind of just narrate what's happening. 00:51:42.240 --> 00:51:44.060 I think that'll give people an interesting sense 00:51:44.260 --> 00:51:45.340 of what this thing is about 00:51:45.360 --> 00:51:46.680 and give us some talking points here. 00:51:47.000 --> 00:51:51.420 So we asked Helix, you know, what plasma samples do we have for the music study? 00:51:51.500 --> 00:51:53.140 And can you break them down by disease groups? 00:51:53.900 --> 00:51:57.840 So to get to that answer, you need to join the clinical data frame with your sample database. 00:51:58.460 --> 00:52:02.100 And then you got to work out which samples are unused, you know, what sample type it is. 00:52:02.460 --> 00:52:05.760 And then for the disease groups, it's got to inspect the clinical data frame and figure 00:52:05.800 --> 00:52:08.040 out how many disease groups do you have in your disease column. 00:52:08.400 --> 00:52:10.220 And then you've got to do the join and the grouping. 00:52:11.320 --> 00:52:14.580 So in this demo, it's gone ahead and done it. 00:52:14.780 --> 00:52:22.680 And it will come back and tell you, well, within our database, we've got like 4,000 samples belonging for Crohn's disease patients and 2,000 with ulcerative colitis. 00:52:23.400 --> 00:52:23.660 Yeah. 00:52:23.840 --> 00:52:28.440 And since people are just listening, let me just go back, just narrate really quick, like fill in a little background visuals. 00:52:29.080 --> 00:52:29.780 You can see it. 00:52:29.980 --> 00:52:33.720 It'll, using the different tools, it'll be like, get sample and it'll run for a second. 00:52:33.780 --> 00:52:36.600 Then it'll write some Python code to do a thing. 00:52:36.720 --> 00:52:40.140 And then it'll do some more data access and then some more code. 00:52:40.300 --> 00:52:46.800 So it's primarily orchestrating a bunch of the tools and the code writing that it already knows, right? 00:52:47.100 --> 00:52:51.040 It's not just reading 100 terabytes of data or whatever. 00:52:51.360 --> 00:52:56.880 No, you can't read all that data because the context window of your AI is limited. 00:52:57.040 --> 00:53:00.660 So you can't just dump all the raw Excel file into the AI and say, 00:53:00.800 --> 00:53:01.600 five video samples. 00:53:03.000 --> 00:53:03.440 Yeah, exactly. 00:53:03.860 --> 00:53:06.620 I mean, even the really big ones have a million context right now. 00:53:06.980 --> 00:53:07.480 All right, carrying on. 00:53:07.780 --> 00:53:09.180 So then it's off to get a picture, right? 00:53:09.480 --> 00:53:12.120 So the next question, we've asked it to do some graphing. 00:53:12.260 --> 00:53:16.360 So we've asked to plot CRP, which is a blood test against cell-free DNA, 00:53:16.660 --> 00:53:19.300 which is a scientific experimental output. 00:53:19.700 --> 00:53:22.840 So to do this, it's got to go and find your clinical data frame 00:53:22.920 --> 00:53:24.420 and join it with your science data. 00:53:24.780 --> 00:53:27.460 And it comes up, writes some matplotlib code 00:53:27.560 --> 00:53:30.480 and gives you a graph back and tell you what it found. 00:53:30.780 --> 00:53:31.960 So that's what it does. 00:53:32.680 --> 00:53:36.160 In this second segment, the AI runs into an error 00:53:36.560 --> 00:53:37.760 and it recovers from the error. 00:53:38.080 --> 00:53:42.500 So the AI is able to read the output of the tool, correct, it's working and come back to you. 00:53:42.680 --> 00:53:46.620 So that's just a demonstration of how an agentic AI system looks like. 00:53:46.800 --> 00:53:47.880 Yeah, good narration. 00:53:48.380 --> 00:53:49.680 And people can go and play that for themselves. 00:53:49.840 --> 00:53:55.240 But I think that's the big difference between what you're saying, just the chatbot as kind of a better Google, right? 00:53:55.700 --> 00:53:57.600 It's not that much of the AI thinking. 00:53:57.800 --> 00:54:00.560 It's a whole bunch of the AI using the tools. 00:54:00.840 --> 00:54:02.480 And the tools are deterministic, right? 00:54:03.120 --> 00:54:04.220 The tools are deterministic. 00:54:04.400 --> 00:54:09.000 And actually designing the tools is one of the most challenging things to do is like, 00:54:09.130 --> 00:54:10.520 how many tools do you expose? 00:54:10.800 --> 00:54:11.740 What should each tool do? 00:54:11.870 --> 00:54:13.560 And what's a logical set of tools? 00:54:13.990 --> 00:54:16.380 And how do you make sure that the AI picks them correctly? 00:54:16.640 --> 00:54:21.880 So I think tool design itself is a huge topic of how you do it. 00:54:22.000 --> 00:54:26.660 But because the tools run on the backend, we can then enforce the RBAC controls on the 00:54:26.820 --> 00:54:27.280 tools itself. 00:54:27.580 --> 00:54:33.300 So you can guarantee that your data access is, you know, security around it is solid. 00:54:33.800 --> 00:54:36.420 Have you thought about adding an MCP server to it? 00:54:36.630 --> 00:54:40.240 So then people can just within their own AI or whatever, 00:54:40.410 --> 00:54:42.520 just, hey, what does Foundry say about this? 00:54:42.720 --> 00:54:44.020 Yeah, MCP is interesting. 00:54:44.170 --> 00:54:46.900 But I think one of the limiting things for us using MCP 00:54:47.000 --> 00:54:48.100 is the governance around it, 00:54:48.270 --> 00:54:51.820 in the sense that we are not allowed to just put data 00:54:52.670 --> 00:54:54.700 into a Claude Code

and send it across to Anthropic. 00:54:55.680 --> 00:54:57.200 Whereas within this system, 00:54:57.520 --> 00:54:59.760 all inference happens in Microsoft Azure 00:55:00.040 --> 00:55:01.280 and it's GDPR compliant, 00:55:01.640 --> 00:55:05.680 which in the UK is important for us from a research perspective. 00:55:06.300 --> 00:55:07.400 So in this... 00:55:07.420 --> 00:55:08.160 It's also important. 00:55:08.340 --> 00:55:10.000 It applies to the US companies as well, 00:55:10.080 --> 00:55:11.580 if they have European customers. 00:55:12.440 --> 00:55:15.040 Yeah, so because of regulatory compliance things, 00:55:15.280 --> 00:55:18.000 we need to make sure that we know where the inference is running 00:55:18.420 --> 00:55:20.780 and we can guarantee that inference all runs within us, 00:55:21.060 --> 00:55:23.220 you know, because we have got the platform controls of Azure, 00:55:23.340 --> 00:55:25.380 we can guarantee where everything happens. 00:55:26.160 --> 00:55:29.060 And that allows us to actually let the AI safely operate on our data. 00:55:29.380 --> 00:55:29.660 I see. 00:55:29.840 --> 00:55:34.320 So maybe you're using a data center in Ireland for your data so it doesn't leave the UK or 00:55:34.390 --> 00:55:34.760 something like that. 00:55:35.050 --> 00:55:37.600 Yeah, we run most of the things in the UK South data center. 00:55:38.420 --> 00:55:41.280 For EU compliance, we run it in the Sweden data center. 00:55:41.760 --> 00:55:43.840 So that's where everything happens at the moment. 00:55:44.680 --> 00:55:45.580 Yeah, that makes a lot of sense. 00:55:46.020 --> 00:55:47.920 Let's zoom out a little bit. 00:55:48.240 --> 00:55:51.180 You've been working on this project for six plus years now. 00:55:51.520 --> 00:55:56.200 You've taken it from working with books to write the Django app to integrating to the 00:55:56.160 --> 00:56:03.960 cloud, Microsoft Foundry, and Azure, and all this tool using AI. What do you see for research, 00:56:04.600 --> 00:56:09.240 either happening now or in the next couple of years with all this kind of stuff coming along? 00:56:09.480 --> 00:56:14.680 Yeah, I mean, I think, you know, we're all very excited about the potential of agentic AI to 00:56:14.900 --> 00:56:20.020 accelerate a lot of the research workflows that we see. And, you know, hopefully, you know, we get 00:56:20.040 --> 00:56:26.760 to discoveries faster because drug discovery is a very long process. And I think agentic AI has 00:56:27.100 --> 00:56:35.360 absolutely a role to play in shortening some of these bottlenecks that we face. So I think it'd 00:56:35.360 --> 00:56:41.240 be very exciting to see. I think agentic AI is going to transform how we do science, how we operate 00:56:41.480 --> 00:56:46.860 also in the clinical world. So I see a lot of potential, but the real world is going to take 00:56:46.880 --> 00:56:52.240 a while to catch up to where the capabilities are today. So I think most people on this podcast 00:56:52.560 --> 00:56:57.960 would have used coding agents, but I can tell you that almost everybody else in the population has 00:56:58.120 --> 00:57:03.780 never touched a coding agent. So the gap between what agentic AI systems can do and what everybody's 00:57:03.920 --> 00:57:09.100 understanding of AI is still very wide. And there's still a lot more work to do to teach people how to 00:57:09.180 --> 00:57:13.820 use the systems, what they're capable of, where the limitations are, and how do you apply it to your 00:57:13.760 --> 00:57:18.860 work. Do you think it can be reliable? Like, do you think we can trust the results and answers 00:57:18.970 --> 00:57:24.080 we're getting from things like Helix and other tools that are working on us? Well, hopefully, 00:57:24.400 --> 00:57:29.680 as things improve over the coming years, the reliability will go up. I mean, personally, 00:57:30.050 --> 00:57:34.660 you know, the coding agents are still not 100% reliable. You know, even like the best models, 00:57:34.850 --> 00:57:40.760 like Opus and Fable, they still can make mistakes. So we're not quite past the reliability threshold 00:57:40.780 --> 00:57:45.980 yet. So we do have to be aware. And I think understanding what you're trying to do is 00:57:46.120 --> 00:57:50.720 actually very important today compared to, say, five years ago, because you have to really 00:57:50.920 --> 00:57:55.020 understand what you're trying to do to be able to supervise an AI to do what you were going to do. 00:57:55.280 --> 00:57:59.340 I totally agree with what you said. But the alternative is to have a human do it. 00:57:59.680 --> 00:58:05.560 And humans are also not 100% reliable. There's recently been this dust up with the Linus Torvalds 00:58:05.580 --> 00:58:12.300 over about using, I forgot the name. There was some AI that they're using as a pre-screen for PRs 00:58:12.360 --> 00:58:16.420 for the Linux kernel. And some people were like, we're not using it. And it's like, look, this thing 00:58:16.420 --> 00:58:22.420 is at least as good as the people often doing it. And it's an accelerator. And there's that tension 00:58:22.600 --> 00:58:27.820 of people expect, I think because it's a computer, people expect it to be perfect, right? Because 00:58:27.880 --> 00:58:33.100 software is typically deterministic. So if it works once, it's always going to work. And AI isn't like 00:58:33.120 --> 00:58:37.380 that, but people also aren't like that. Yeah, that's right. I mean, like, you know, 00:58:39.000 --> 00:58:43.080 certainly everybody, you know, having an AI by your side is like having a colleague on your team, 00:58:43.400 --> 00:58:47.660 except this colleague can write code like tremendously faster and much better than most 00:58:47.820 --> 00:58:53.460 people. You know, today I would say that Claude Code and codex can write code way faster and way 00:58:53.660 --> 00:58:59.180 better than me, but you still need that kind of strategic view from the human to

just make sure 00:58:59.200 --> 00:59:00.640 that you're going on the right path 00:59:00.940 --> 00:59:02.520 because you got to drive the AI 00:59:02.600 --> 00:59:03.600 and you've got to direct it. 00:59:04.530 --> 00:59:05.920 And it's the same with the scientific work. 00:59:06.660 --> 00:59:08.000 So I tell my scientific colleagues, 00:59:08.090 --> 00:59:09.800 you need to make sure that the answers 00:59:09.980 --> 00:59:11.580 that come back pass the smell test. 00:59:11.650 --> 00:59:12.520 You need to make sure that, 00:59:13.300 --> 00:59:14.440 you know, the numbers that you're seeing 00:59:14.510 --> 00:59:15.980 are in line with your expectations. 00:59:16.280 --> 00:59:17.000 You're expecting a number 00:59:17.110 --> 00:59:18.080 in this magnitude range 00:59:18.150 --> 00:59:18.820 and you get it there 00:59:19.160 --> 00:59:20.480 because sometimes the AI goes off 00:59:20.480 --> 00:59:21.460 and does funny things, right? 00:59:21.720 --> 00:59:23.000 Yeah, I just, you know, 00:59:23.140 --> 00:59:24.600 if you ask it to verify everything 00:59:24.880 --> 00:59:26.940 and prove it and write code to back it up, 00:59:27.090 --> 00:59:29.160 like it's better than if you just ask it 00:59:29.180 --> 00:59:29.780 You know what I mean? 00:59:29.940 --> 00:59:31.500 Like there's techniques, 00:59:31.730 --> 00:59:32.700 but you do have to treat it 00:59:32.780 --> 00:59:33.820 with a little bit of skepticism. 00:59:34.220 --> 00:59:35.540 But that's also true for your colleagues 00:59:35.960 --> 00:59:38.140 and your grad students and whatever, right? 00:59:38.400 --> 00:59:40.440 Like no professor would just take a, 00:59:40.770 --> 00:59:42.560 like, hey, grad student, write the paper. 00:59:42.700 --> 00:59:43.440 And they don't even read it. 00:59:43.440 --> 00:59:43.920 They just publish, 00:59:44.090 --> 00:59:46.160 they just send it off to nature or medicine 00:59:46.290 --> 00:59:48.960 or write the medical journal or whatever. 00:59:49.400 --> 00:59:50.100 I mean, it depends. 00:59:50.230 --> 00:59:51.520 I think for disposable stuff, 00:59:51.680 --> 00:59:53.600 you can let the AI do more of that. 00:59:53.780 --> 00:59:55.300 But for the real critical workflows, 00:59:55.540 --> 00:59:57.260 you want to make sure every single step is correct. 00:59:57.580 --> 00:59:57.680 Yeah. 00:59:58.080 --> 00:59:58.180 All right. 00:59:58.260 --> 01:00:00.780 Well, that brings us to our final call to action. 01:00:01.420 --> 01:00:05.440 If scientific researchers or medical researchers are out there listening, 01:00:06.100 --> 01:00:07.800 either what can they learn from your experience 01:00:08.180 --> 01:00:11.860 or if they wanted to work with you on some of this, what would you say? 01:00:12.180 --> 01:00:15.220 So one of the big questions and why we're putting Foundry out there 01:00:15.220 --> 01:00:16.940 is we don't know how generalizable it is 01:00:17.340 --> 01:00:21.520 or whether it's just going to be hyper-personal software for ourselves, our team. 01:00:21.700 --> 01:00:25.000 But I think the general principles, I've shared them widely. 01:00:25.320 --> 01:00:28.140 So please feel free to just take the idea and run with it. 01:00:28.240 --> 01:00:31.500 I'm very excited to see what other people can do with the ideas. 01:00:32.150 --> 01:00:35.420 And perhaps within their own institutions, they might want to build their own systems. 01:00:35.670 --> 01:00:36.780 And I think that's absolutely valid. 01:00:37.440 --> 01:00:42.160 But if you do want to explore a partnership with us, come and visit us on foundry120.com 01:00:42.190 --> 01:00:42.940 and drop me an email. 01:00:43.620 --> 01:00:43.940 Yeah, great. 01:00:44.080 --> 01:00:46.260 And I'll link to your web page. 01:00:46.430 --> 01:00:49.740 You've got email and your GitHub and other ways to get in touch with you there. 01:00:49.890 --> 01:00:53.240 You also have a lot of interesting writing here, so people can check that out. 01:00:53.520 --> 01:00:56.160 Yeah, that was my journey into coding, really. 01:00:57.420 --> 01:01:00.020 Yeah, we all have one of those journeys to tell the story of. 01:01:00.450 --> 01:01:02.140 Well, Sean, thank you so much for being on the show. 01:01:02.620 --> 01:01:03.260 Keep up the good work. 01:01:03.350 --> 01:01:04.840 I think this is a super interesting project. 01:01:05.060 --> 01:01:05.740 Thanks very much, Michael. 01:01:06.110 --> 01:01:08.340 Very happy to be here, and thanks for the invitation. 01:01:08.740 --> 01:01:09.380 Yeah, you bet. Bye. 01:01:09.540 --> 01:01:09.660 Bye. 01:01:11.099 --> 01:01:13.480 This has been another episode of Talk Python To Me. 01:01:13.860 --> 01:01:14.580 Thank you to our sponsors. 01:01:14.810 --> 01:01:16.120 Be sure to check out what they're offering. 01:01:16.290 --> 01:01:17.640 It really helps support the show. 01:01:18.420 --> 01:01:19.960 This episode is brought to you by Sentry. 01:01:20.760 --> 01:01:23.760 You know Sentry for the air monitoring, but they now have logs too. 01:01:24.050 --> 01:01:26.780 And with Sentry, your logs become way more usable. 01:01:27.200 --> 01:01:31.140 interleaving into your error reports to enhance debugging and understanding. 01:01:31.760 --> 01:01:34.880 Get started today at talkpython.fm/sentry. 01:01:35.500 --> 01:01:38.380 And it's also brought to you by Talk Python Courses. 01:01:38.920 --> 01:01:41.220 Course completion certificates are now live. 01:01:41.400 --> 01:01:45.640 If you finished a course, there's a certificate waiting for you on your account page right now. 01:01:46.140 --> 01:01:49.720 Download it as a PDF or add it to your LinkedIn profile with one click 01:01:50.310 --> 01:01:51.860 under licenses and certifications. 01:01:52.560 --> 01:01:53.740 Same section as your formal degrees. 01:01:54.960 --> 01:01:58.800 Visit training.Talk Python.fm slash account to see what you've already earned. 01:01:59.600 --> 01:02:02.920 And if you're not already subscribed to the show on your favorite podcast player, 01:02:03.580 --> 01:02:04.220 what are you waiting for? 01:02:05.000 --> 01:02:06.660 Just search for Python in your podcast player. 01:02:06.760 --> 01:02:07.640 We should be right at the top. 01:02:08.060 --> 01:02:10.900 If you enjoyed that geeky rap song, you can download the full track. 01:02:11.040 --> 01:02:12.880 The link is actually in your podcast blur show notes. 01:02:13.780 --> 01:02:15.100 This is your host, Michael Kennedy. 01:02:15.400 --> 01:02:16.560 Thank you so much for listening. 01:02:16.780 --> 01:02:17.540 I really appreciate it. 01:02:18.000 --> 01:02:18.700 I'll see you next time. 01:02:30.340 --> 01:02:33.140 I'm out.