HIMSSCast: AI-generated code needs AI-powered review systems

Speaker A: When an AI makes a mistake in coding, they're not uniformly, uh, random. It's not like, oh, I made this particular kind of mistake that I've never made before. They make telltale mistakes.

Speaker B: Hello and welcome to himscast. I'm Jessica Hagan, executive editor of Moby Health News. And today on the podcast we have Redesign Health's head of technology, Aaron Santo Oram. Thank you so much for joining me.

Speaker A: It's wonderful to be here, Jessica.

Speaker B: It's wonderful to have you. So, uh, can you tell our listeners a little bit about Redesign Health?

Speaker A: Absolutely. We are a, uh, venture and applied technology company focused on building the next generation of healthcare. Uh, through our investing work, we manage two strategies. The first is our early stage venture capital work where we back the world's most talented venture founders who are building these transformative companies in our portfolio. The second is our venture buyouts or our later stage strategy where we partner with very talented management teams who run companies that have already earned pretty substantial scale and help them reimagine and rewire their core systems around AI. We've been around since about 2018. Uh, since then we and our portfolio companies had the pleasure of, uh, touching the lives of more than 15 million patients here in the US and around the world. Uh, we operate in almost every area of healthcare, we like to say, from pediatrics to senior care, everything in between. And through the course of our work, we've had the very good fortune of working with, uh, leading health systems like Mayo Clinic, upmc, Cedars Sinai to help imagine and build towards the next generation of healthcare delivery.

Speaker B: Right. And what's very interesting to me is that Redesign Health recently created its own AI that is checking, uh, for code that's written by AI. Right. So like, AI has gotten pretty good at writing code, but it is not so good at checking that code that it writes for mistakes. So your company created argus, which, uh, is an AI powered code review system designed to review code specifically written by AI. Right. Can you tell me more about argus?

Speaker A: Absolutely. So one of the key functions of our engineering team is helping redesign itself, practice as an AI native organization. And what we found is that this isn't just giving people better access to models or use this tool or use this other tool. Uh, it means enabling every domain expert at redesign, whether they sit in legal, investing, operations, finance, clinical strategy, to turn their expertise into software and systems just about as quickly as they can think through the problem and the solution. And um, you think about it, this is a pretty crazy shift because AI has deflated the cost of creating software so, so fast that people without traditional engineering backgrounds can now create genuinely useful systems. Uh, but if you've vibe coded anything yourself, or if you uh, read about how other organizations are bringing AI to bear on their work, you know that you can get to a proof of concept pretty quickly. But the real important problem is figuring out how you can make the system performant and well governed and safe and fast. And that turns out to be really challenging. ARGUS is a system that responds to the uh, fact that you can create just unimaginable amounts of code. But to put AI in the hands of domain experts and to ask them to create end to end scalable, really useful systems requires that we take what used to be code review, which engineers used to do for the entire organization, and put that capability into the hands of people who don't have this traditional engineering background. So we've thought through how to take any kind of code, whether written by, uh, an engineer with 40 years of coding experience or someone who started coding four days or four weeks ago, and uh, imbue it with the quality and safety good governance that all of our software at redesign around our portfolio need to uh, bring to bear in order to be successful. So argus, you can think about it as kind of a quality control and code review system. I should say we're not the first to come up with this. There are a lot of organizations that are working on agentic code review. We tried them all and basically found that none of them worked as well as we needed them, uh, to, for our particular set, uh, of use cases where engineers and non engineers alike are developing and shipping really high quality technology. So we ended up realizing that we have to build our own system for the problem.

Speaker B: I feel like it sounds like, okay, maybe this isn't a difficult thing to build, but really it's incredibly complicated. Right, because like, if you're thinking about it, you, you could, there could be so many bugs within the AI because there's like all these different APIs that people implement and there's all of these different, if like somebody does like a migrate, you know, migrates their database to like some, it just seems like there would be so many options, so many like, circumstances where it could just go wrong.

Speaker A: Yeah, absolutely. And so one of the things that we've had to think through is when an AI makes a mistake in coding, they're not uniformly random. It's not like, oh, I made this particular kind of mistake that I've never made before. They make telltale mistakes. And so we've built Argus to work through in an almost expert and specialist way the different types of mistakes that AI tends to make. So as an example, when someone creates some software that we run through Argus, ARGUS will break down the entire system into its individual flows. Here's what I show the user on the front end here is the flow, uh, of the data through the system. Here's the permission structure, here's who's allowed to see each piece of data. And then finally, what is the big risk? Are we making investment decisions on this? Are we sending out emails to founders or prospective founders? How important is it that we get this exactly right? And then we'll basically dispatch these highly specialized reviewers, uh, to uh, work on individual small components of this, for example just data security or just performance or just uh, data governance or just front end performance. And when you stitch all of that together and then you put another powerful AI reviewing module on top, what comes out is trying to, is software that is much, much less likely to make the telltale mistakes that AI tends to make because we've kind of, uh, specifically tried to orient the system to help catch them. And uh, the way that the system improves through time is that when it catches a mistake, when it uh, makes a change, when it suggests a different way of doing something to an engineer or to someone who's building the system, then it's able to kind of feed the result back into uh, its own training set. And so as we use the system, as we develop more and more technology, as we impart more and more expertise into the system, the uh, ARGUS gets better and better each time that we use it.

Speaker B: So what about hallucinations? It sounds like there's always going to be somebody who's looking at whatever is coming out of Argus, right, to double check it. So it's more of a tool as opposed to a complete solution that's always going to find something.

Speaker A: Well, what I would say is that Argus is kind of the uh, quality control layer so that whatever you put into it, what will come out is something that we either, uh, can be uh, confident is a very high quality and ready to go to production or something that we need to spend a little bit more time thinking about. And more often than not, when something doesn't pass the ARGUS review and we can't get it to a place where it does, there's something more fundamental than just the software that's wrong. For example, you might be trying to expose data to a type of person who's not supposed to see that data. But Argus is 100% agentic and fully hands off. So the way that it works, ah, is you put software in, uh, Argus will, uh, figure out what the issues are with it, and it'll feed those results, uh, back into your coding system, which will then make the changes and lather, rinse, repeat, so that what comes out of, uh, the many different cycles or iterations of Argus is something that is actually just ready to go.

Speaker B: How long did it take you to develop this?

Speaker A: Uh, we built this over the course of about two months. Um, but you know, as with everything in AI, I believe that this would have taken two years, two years ago, and we could do it in two weeks now. And we fully imagine that the problem of agentic code review will be something that is much more substantially solved in the near future. So as with everything, we're just getting used to building something really complicated and then being very okay, throwing it away as the frontier continues to shift. But in this instance, we found that we had to build our own system because nothing that was off the shelf was actually of high enough quality.

Speaker B: So that's what I was going to ask you. So as technology, technology, especially AI, I mean, it's like every single month you hear about a new type of AI that's coming out. Right. But like, how are you making sure, or how are you going to progress the technology as more, as new type of emerging technologies are developed?

Speaker A: Yeah. One thing that we think constantly about is where do we have a right to play? We are not training our own foundation models. And so how do we develop our systems to just get better and better and better as the frontier of intelligence, as these language models get more and more powerful, when opus 4.9 comes out, when e 5.6 comes out, how do we make sure that our systems are designed to constantly improve? And so the, uh, the environment, uh, that we build around language, uh, models called a harness. These are the skills, the data sets, the flows, the human reviews and policies that we build in. And so we take care at the beginning of every project to think about whether the system gets better with more intelligence. And if the system gets, uh, better with more intelligence, how do we ensure that it kind of stays ahead of the frontier each time?

Speaker B: Right, Absolutely fascinating. Aram, um, thank you so much for joining me. I appreciate your time.

Speaker A: Wonderful to be here. Thank you again, Jessica.

Speaker B: And then to our audience, thank you for joining us as well. Just as a reminder, if you like what you hear, please subscribe to Hymns cast on Apple Podcasts, Spotify or Amazon Music.