Previously Impossible

Host Intro (Gene Kim)

All right. So our first talk is Owen Parker, CTO, and Kurtis Cochran, CFO of Arkana Laboratories.

So earlier this year, thanks to an amazing person named Sadie Clark, who is in the audience, I got to learn about the amazing work being done at Arkana Labs. So they are the country's leading kidney pathology lab. Renal pathology is the clinical term, but what it means is when something goes wrong with your kidneys, they're the ones trying to figure out why. And so if you get an abnormal kidney result, there's a good chance it goes to them, or maybe perhaps the number two player in the market, the Mayo Clinic.

So I got to learn about the work of Owen Parker, the CTO, and how he's utterly transformed the organization. And so the next thing I know, I'm visiting them in Little Rock with my buddy Steve Yegge, and there were so many highlights. One of them was meeting Dr. Sri Sharma, one of their owners and their director of lab services, who assembled a group of 25 researchers from the academic research community. And so Sri and I got to show them how to vibe code to do data analysis and help them generate potentially publication-level quality research papers that otherwise would require a grad student for a month.

So the other highlight of my day, among many others, was meeting their CFO, Kurtis Cochran. We had a big team dinner that night, which was a lot of fun, but I think when people heard Kurtis and me talking into the night about the horror stories of spreadsheets and general ledgers, everyone else left. And I wasn't offended much. But for me, it was amazing to hear what they're doing, especially from a CFO's perspective.

So here to tell that story is Owen and Kurtis.

Owen Parker

Thank you, Gene. Very kind. So I left Kurtis backstage for just a moment because we're going to talk about some of the actual numbers, not just the numbers I gave him. But he'll be out here in just a second.

Let me tell you, being at an event like this is so much better than trying to benchmark against peers in healthcare. I finally feel like I'm not a crazy person. So thank you all for your participation in this.

Got that backwards.

So just to kind of set the stage and present the stakes: about 37 million US adults have some form of kidney disease, chronic or otherwise. And for all of the biopsies that come through Arkana Labs on a daily basis, these are kind of the rare and definitely chronic forms that most of the time have zero cure and zero treatment or zero therapeutic.

For a failed kidney, you can perform a kidney transplant that will replace it. Dialysis helps filter when a kidney is no longer capable of doing that. But ultimately, neither of those things truly restores kidney function in any meaningful way. So two things are at stake in the work that we do. One is finding more effective therapeutics and treatments, and two is getting it to the people that need it quickly.

This is the mission statement for Arkana Labs, and again, I'm Owen Parker, the CTO there. So the way our workflow operates is a biopsy is performed — typically it's ordered by a nephrologist, a kidney doctor. That biopsy is performed and sent to us in a package overnight from 48 states. And what we do is we process that through our laboratory starting at five AM the following day, and our pathologist will create a diagnostic report that very same day. So you have biopsy today, you have result tomorrow, and that's incredibly important for people that are struggling and are in a time where they're waiting on some kind of answer about what's going on. The treatment plan is very heavily influenced by the diagnostic report that our physicians generate.

So Gene alluded to this, but the scale of Arkana Labs: we do well over half the kidney biopsies in the United States, and we're only about 250 people. So 30 of those are in the technology org, but that's not really what the talk's about. The talk is about what we've been able to achieve in the last 18 to 24 months, which is staggering for a healthcare org this size with this amount of people.

I had an analogy that I was super excited about, but then Nick Kirsten stole my thunder a little bit, so I'll kind of alter it some. Really, what I wanted to drive home with this is that with every single one of these transportation revolutions, you sort of had trade-offs that existed. With a horse, you had maximum optionality — you didn't need the roads, you didn't need the trails, but you couldn't go that far, and you couldn't carry very much. The train was sort of the polar opposite of that, where you could carry a whole lot and go really far, but you could only operate on the tracks that were given. And then you sort of had the car and truck, which clawed back elements of that optionality but didn't really ever deliver fully on either one of the train or horse's promises.

What we have found at Arkana, in our experience in the digital workflows — we have many atom-based employees still that are doing the things: handling tissue, slicing, doing staining. But in our digital-based roles, what we have found is you get the horse, the car, the truck, and the train all at once, right? There isn't this trade-off scenario anymore. You don't have to choose optionality over throughput.

So this is what I conjured up in my head when I thought about this. And honestly, I made this before any of the news broke over the last couple of weeks. But if some of these people at Frontier Labs were extrapolating out their work and saw this in the future, I can understand why they wanted to get out there and kind of spread the word on it. But ultimately, this looks pretty cool — I would probably drive it. I don't know if you drive it, but...

So again, the mistake is thinking that you still have to pick when you can have your Frankenhauler, and you can have it the way that you want it.

So this is kind of how we frame the different categories of AI at Arkana. This first bucket of efficiency AI — these are the things like summarize meeting notes, draft an email for you, create a slide deck like this slide deck. These are valuable things, but ultimately they're table stakes. Everybody has access to them. There's nothing about them that gives you any kind of advantage. The second layer of this is the tearing up the tracks element. This is rebuilding a core process and putting AI at the center of it instead of just sprinkling it around the outsides, which we see all too frequently. And then the last one is this greenfield element. This is building the routes that nobody thought were possible. So this is the previously impossible component, which is the title of our talk.

And really, "previously impossible" is something that we landed on because I think it really describes the experience that many of our — especially non-technical — stakeholders at Arkana have. It encapsulates what they're feeling.

I would say directionally at Arkana, we've seen probably far less than 10% of our value come from that efficiency bucket. We've seen a lot more come from the rebuilding, and we've seen even more come from the greenfield, even though it's a little bit harder to put your finger on what the ROI is.

So basic claims for this talk. One, efficiency AI is table stakes. We've seen a whole lot more value through the other two categories of rebuilding and greenfield. Being small and regulated is something that for us has turned out to be a huge advantage. We all know this as technologists — constraints force innovation, and so when you work in an industry that is as heavily regulated as healthcare, sometimes it's more compliant and it's faster, and it's more capable to build internally and to bring that function in-house. And then lastly, the greenfield compounds. This is the thing that is the most exciting because ultimately, when you own the systems and you build the systems, you're curating the external and the internal products, and those things continue to build a competitive moat and expand that moat.

So in the last two years — probably 18 months more specifically — this is just a list of some of the things that we have accomplished. You don't have to know what any of these mean. The thing that I really want to highlight is this bottom red-highlighted one, which is why it's red. We did a lot of this from inception all the way through to delivery in most cases with an engineering, product, and software team of fewer than six people. So these are all core processes for our business, and our team has expanded past that and our capability set has expanded past that pretty dramatically, but fewer than six people were able to achieve this.

And one of the places that this landed really quickly was actually in our finance department. And so this first box on the top left — our CFO, Kurtis Cochran, is going to come and talk through right now.

Kurtis Cochran

Morning, everyone. It's good to be here.

I'm Kurtis, CFO of Arkana, and we haven't had a lot of finance guys on stage at this conference yet. My biggest takeaway from yesterday was the CEO of Highland — if you caught it — he said that he could never trust a CFO that used a Mac. Owen's been trying to get me on a Mac for a long time. I think with some of the vibe coding, it probably makes sense for me to have a Mac. He'll have to support both, though.

So this morning, I want to just quickly illustrate something that we did in my organization that was an orphan process and often overlooked, and that was in my medical billing department. A few of you probably just fell asleep. I know medical billing is not riveting, but I think maybe this particular illustration you may learn something from.

So medical billing is really complex and tedious. At Arkana Labs, we get about 150 biopsies every single day. Each one of those biopsies generates hundreds of lines of itemized billing code that then has to be perfect before it goes out the door, or someone who's supposed to pay me is going to find an excuse not to. So it's really tedious and important work.

But I do want to get a little bit into the weeds here because I think it'll illustrate some things. So currently, my billing team works out of two primary systems: our lab information system, LIS, and our billing system. And so when a biopsy comes through our door at Arkana, we put our patient's information into the LIS system, and then as it goes through the process, we track what we did and why we did it, and we augment that patient's account with important data that then gets translated into something called an HL7 message. That HL7 message is a universal language that billing systems can read and turn into an invoice. Underneath it, it's just structured language, which is an important point I want to come back to.

So under our old way of doing things, before we came up with a solution, this message would get sent at the end of the day with all of the work that we did into my billing system and then translated into an invoice — which was the rub for our business, because we have a unique way that we do our tests. We do a panel of stains, they're called, and each stain we have multiple duplicates of, and then each one of those stains has to be coded differently to our payers. And when that message, that HL7 message, would go from the LIS to the billing system that we use, it would get translated wrong. And so every single day, my team has to go into the billing system, open up every single invoice, and manually correct thousands of entries into this billing system to be sure that I can keep my doors open at Arkana.

So this is a huge problem. Like I said, it's tedious but really important work, and it needed to be done.

Then came along AI, and with the help of Owen's teams — that he'll talk about a little bit today — and Jean Kim, my head of that team, Becky Schmitz, in just a matter of weeks, starting from knowing nothing, was able to learn how to build something using AI.

And now going back to the point that it's just structured language — it's structured text — what Becky was able to figure out is that we're able to read that text and come up with a solution before it gets to the billing system, and we don't have to wait on our vendor to do that. And we had talked to our vendor many times through the years asking them to solve this for us, but we would be the only people that were going to use that code, so it just didn't make any economical sense. It was impossible in a sense. We weren't going to make the coding investment. They weren't going to make the coding investment. So we just had a manual solution to something that should have been digitized a long time ago.

So AI comes around. Becky's able to learn how to build this tool. And so now what we've done is we intercept that message from our LIS. We read the HL7 message, we apply the same logic and rules that my team were already doing further down the value chain, rewrite that HL7 message, and then send it into our billing system, which now properly reads that message. And what do I get? I get an invoice with no errors relative to our staining codes. Huge improvement.

What did I ultimately get out of this? No staining errors, so no duplication of work. I save a full FTE, and that's about 10% of the team, and there's always more work to do. We ignore small little bills here and there all the time because we don't have the time to mess with them. So now we have more time to mess with bills that we didn't previously have the time to mess with. And we have a platform that we can continue to build some agentic processes on top of.

So other than the story of doing something in medical billing — very boring area — that was previously impossible, I have two other takeaways for you on this.

One is the operational frontier was the one that actually came up with the solution, the subject matter expert. We've heard a lot about that in the last day. Don't be afraid to let your subject matter experts come up with solutions along the way. And I have a bit of a sub-point on this, because I'm a CFO, so I think this could be valuable for the crowd: all of us are sitting here thinking the spice must flow — do not turn off my tokens. Okay? I'm the one that's got to pay for those tokens.

So what you could do inside your organization is help your finance team understand how to account for these tokens, because they're sitting there with their visors, counting beans, asking you, "Show me what the ROI is. Help me understand how I'm getting something out of what I'm spending on this." Help them understand where to actually account for that. Does it belong in IT? Does it belong in, in my case, the medical billing system's ledger? Because when you put this in the right spot, it makes it easier for everyone to understand what the return is on the work that's being done. You can see that I've saved an FTE for X amount of tokens that I spent to both develop this product and continue to maintain it over time.

The second thing is that this gave me a ton of leverage over my vendors. Another ROI point is: also help your teams understand that when you can build something that previously a SaaS solution could do, it gives you leverage over them. So you can find token money by getting cheaper software. Negotiate your contracts knowing that you can replace them, or some portion of what they do, or help them understand they're not providing the same value that they used to, and ask them to double down on what they were doing for you.

So with that, Owen.

Owen Parker

Yeah. Thank you, Kurtis. So again, we saw a lot of this value land outside of the technology team to start with. And for my team, where it landed first was in the Arkana Connect app. This is the primary interface for delivering a diagnostic report to a nephrologist, which is ultimately our customer. We're serving patients at the end of the day, but our primary customer that we actually deliver that report to, to help inform the treatment plan, is the nephrologist. And this was an app that really probably accounted for maybe 10% of our total delivery of diagnostic reports. We were still doing faxes. We were still doing all of the really cool things that we were also doing 25 years ago.

And what we did is we had farmed out a rebuild of this to an external vendor, and shockingly, they were both over budget and behind schedule in a really meaningful way. And so we sort of had this powwow internally and decided that, you know what, we probably can point these resources inward and hire our own team and deliver something that is ultimately better. And so we did exactly that.

And here are some of the metrics. We saw engagement on that application go up 5X — we're well over that at this point — but we saw all of the stability metrics increase. We've been able to add some really critical features like clinical trials matching, super important for people that are dealing with rare forms of disease. And so we've been really, really excited. But that's kind of the efficiency play, right? Turning those dollars inward or saving the money is the efficiency play. What we have found is that we now have a software engineering team, which is a capability unlock that didn't exist and doesn't exist with any of our peers in the marketplace.

So this brings me to our most active initiative currently. Kurtis talked about the LIS system, a lab information system. It's no different than any HR system if you've worked in a hospital or medical practice that's not a lab — it's a validated system of record. Every HIPAA-regulated organization has to have one of these to store patient records. Ours is a system called PowerPath, and it is a 25-plus-year-old base system. There's no SaaS offering. There's no cloud offering of note. So this sits on a box in the closet downstairs on the first floor of our Kansas building. And for 20-plus years, we've built a business around this box. We've had to accommodate for its lack of features — and a really wonderful development team on the other side — with things like team messages for case counts, shared Outlook inboxes for special stain ordering which happens throughout the day. All these workarounds that ended up on paper and spreadsheets or clipboards and paper logs — all of those things have been a result of the track that was laid by another vendor.

We, at our scale, especially four or five years ago in a pre-AI era, would have never greenlit a full replacement. It just wouldn't have made sense. The economics aren't there. Nobody's going to build a piece of software for one of one. We're the largest in our industry, so our problems are very unique to us.

So what we started doing is we're sort of eating this piece of software from the outside. We've started with these modules that don't actually have functionality in the core system currently, and we're eating it from the outside. So accessioning is a good example of this. In the morning, with every biopsy, we receive a packet of paperwork — a physical packet of paperwork. And we used to have a person that was thumbing through that and pulling out the relevant details, keying them into PowerPath, and then proceeding throughout the day. That's now a scan, and it hits a bunch of agents. It writes to the correct places. Accuracy, speed — all those things have increased really dramatically, and it's allowed us to kind of kick our process off a little bit faster. We've created a report builder system for our physicians, which is templatized reporting. It's a more consistent product for the nephrologist externally. They may have one pathologist look at their case one day and then another pathologist look the next day — there's going to be nuanced differences in how they deliver information. And so having a really consistent system with a consistent template has made a huge difference in the product quality.

So this is kind of the greenfield component here. We are capturing the data more often and further upstream than we ever have, and that's allowing for that improvement cycle to happen a lot more rapidly. And as we continue to build this out, this is something that a competitor ultimately is not going to be able to copy, because we have way more volume, we have unique processes, we have specific experience and expertise that nobody else has, and we have a huge concentration of it.

So what is the next greenfield? Because it never ends. This is something that we're super excited about, and we feel a great deal of responsibility around. We have started to do open-weights medical foundation model training. We have a corpus of data that consists of about two million physical slides, which is a huge number. We've never thrown away a slide in the 26 years of existence. We also have the paired diagnostic reports. So we have this incredible corpus of data, and we've decided that what we want to do is turn that into something that's more widely available and can aid in the clinical as well as in our partnerships with outside pharma companies.

The thing that we really care about here — these are some fun, fancy statistics. This is off of our training runs two and three. We've obviously had more than that, but I'm not going to share all the details because some of that stuff we're still working through. But the base model was able to achieve the thing that matters to us about 35% of the time. So that is: we put an image in, and what we get out on the other side is an identification of the specific morphology or the tissue elements that matter to the diagnosis, or some kind of narrative description of what the disease might be. And we got that about 35% of the time. After this third training run, we got 85-plus, which is huge. And really the long tail — the 15% gap that we need to close — is related to the really unique and specific disease types that were underrepresented in the datasets. And so this is a really, really exciting thing.

We know that next-token prediction accuracy is not the same as diagnostic accuracy. It is extraordinarily different. But at the end of the day, we have what exists in this black box, which is the biopsy images and the diagnostic reports, and we have them in spades. We know that we have to get more of what led the patient to that encounter — a lot more of their clinical history. We know that we have to understand the treatment path decided on, and then we know that we have to understand the outcome of that treatment path in order to really execute this fully and get it to a point where the FDA might actually take a look at it.

So none of this came from selecting the right tools. This is the people component. So that's the technical — here's the people and process.

Gene mentioned Sadie earlier today, just this morning. She's actually in the audience today, and she is Sadie Blassford now, by the way. But she actually developed this internal cohort that was focused on non-technical stakeholders, took them through a training process. It was actually a curriculum derived from Gene and Sei's vibe coding book. And she's repeated that process and refined it multiple times. And it's had a huge impact on our adoption rates and the ability to solve problems.

We've got real experimentation with direction, which I think is super important. For non-technologists, I have found that there is a massive blank canvas problem. When you present them with a tool and say, "This can solve things for you," people can't figure that out on their own. They really do need to do the experimentation to have the aha, but they need a lot of direction around the experimentation.

So this is also the point that we've tried to make with our people — we're helping them reimagine or redefine the work process. So if you think about a job, it's kind of a collection of tasks. Those tasks can sit in any one of these buckets. And so what AI is doing is it's eating the middle, right? The actual execution is being eaten really, really quickly, and we're seeing the value equation migrate to the edges. But it's not zero-sum, so we're seeing the edges expand as well. We're able to take on additional capabilities, which we're super excited about.

So two things to take away from this. One is: if you're small and you're regulated, go find an engineer. Or if you're an engineer, go find a small regulated company that's got really specific problems — they're super fun to solve for, and there's great opportunity to participate in upside on them. And then if you lead one of those, go find an engineer, right? Stop just taking the track that was laid by other people.

Here's what we need help with. Obviously, we're a growing business. We need the talent, we need the datasets, and we need advisors and collaborators on all of the things that we're taking on right now. So if you know people that live in this space, that breathe this space, please send them my way. I'd love to have a conversation with them. Thank you all. Appreciate it.