The Exchanges

Q Great. So maybe for the final part of this conversation, uh, switching hats again. So you're, you're, you're an investor. Maybe, uh, talk about, um, so what is the fund and what do you guys do? Who else is involved? Uh, just like, you know, quick, uh, a little bit of pitch, I guess.

A Yeah. Quick one. Um, so it started after my first company MetaMind was acquired by Salesforce. Uh, and I, Kind of kept my grad student lifestyle for a few more years after that. Um, and so I had extra funds, and so I started investing in all my smartest students and friends and coworkers and employees and interns and whatnot. Uh, and I guess, like, I was very fortunate because especially in the early days, 20, you know, 1516, 17, like, you had to be pretty smart, pretty motivated, and excited about AI, and also a little bit contrarian Uh, to want to learn about neural networks for me, and it's like writing like the first, you know, like first lecture worldwide on like NL, like neural nets for NLP and stuff. And so I was very lucky to have incredibly smart, uh, people in my community. Uh, like for instance, I invested, uh, in, uh, these two, uh, three founders, uh, creating this cute company called Hugging Face. I had a beautiful five million dollar valuation. Um, and now they're worth, like, four and a half billion, right? And so it worked out pretty well, and after doing this for a couple of years, uh, and having a lot of fun, just like, you know, I had a full-time job, uh, but I love AI. I want to make sure that AI has positive impact on the world. It's general purpose technology, so, you know, it can go in a lot of directions, and so you want to spend more energy on the posi…

AI assessment note: “I started investing in all my smartest students and friends and coworkers and employees”

Q a technical perspective, why is it, why is it possible to do this, um, in the first place? You, you know, you would hear a lot in the market that, like, you, you cannot be the database that does it all. You have to, You know, choose and, uh, you know, make trade-offs. So why, you know, why is this happening? Why are you guys able to do this today?

A Yeah. So the separation of storage from compute is an important part in that. So being able to scale those out depend, you know, dependent on what you need in your application. The biggest change in CyrilDB is how we deal with the IDs of documents. So in Postgres or relational database, something like MySQL, you traditionally have an ID field which is an integer. So one, two, three, four, five, six, seven, eight, nine, 10, and so on. And you can change that, but that's, that's pretty much how it's initially formed. Um, in something like MongoDB you have an object ID, and that is a A string, for instance, and then graph databases are slightly different, but in CyrilDB, the ID is made up of the table name that the record belongs in, and the ID itself. So that ID can be a number, it can be a string, it can be a UUID, uh, and it can be something more complex, which I'll come to in a minute. Um, by Changing that one small piece of technology enables you to bring all these models together. So, if I want to do very quick requests to get just one record from the data store, maybe I'm in, I'm in ads and I need very quick response time for a very particular record, I can go and get person, Toby, and I don't have to use an index, I don't have to know anything else about the data, I can just go and pull that record immediately from the store without having to access anything else in the ta…

AI assessment note: “Changing that one small piece of technology enables you to bring all these models together.”

Q Because we're kind of early to the whole wave of, um, Building, um, great database companies on, on top of Postgres. Is that, is that fair or no?

A I think we were possibly the first company to really do an extension. I mean, Postgres has been a platform for a while, but in the past, most people have forked the database. Uh, Pivotal does Greenplum. Amazon Redshift is actually a fork of Postgres eight. Uh, they generally, they dead end then. And the problem is they diverge from the community, which is one of the largest You know, companies out there, or one of the largest data, data platforms out there. I think the other thing is that we knew from the beginning that we were also building a managed cloud platform, and so people in the past would think of extensions as like tiny little plugins, operational things, and really, really were probably one of the first to think of it as taking it into new places where it hasn't been before.

AI assessment note: “I think we were possibly the first company to really do an extension.”

Q And, uh, so you have all of that, and then you, um, funnel it, uh, into the platform, or I guess, I guess maybe you do all of this once it's in, into the platform already? And you said the platform is, uh, is Google for, for that, for, um.

A Yeah, ZDP is built on Google, BigQuery right now. So, you know, all these pipes are that are continuously pushing the data to the platform, or sometimes we are pulling the data, then you have to definitely store it on, on a platform like that, and then you go to the processing on top of it, right? So you have to sometimes join these, this data, extract certain attributes for each of these, create a graph, because these are related to that, Uh, related to each other, you know, a person works in a company, or this company maybe sells a product to this another company. All these connections are there that we are extracting out of that, and that is done through, you know, different technologies, you know, either spark code or data flow that GCP has, and all kinds of basic processing. Machine learning, definitely, right? Model building, some of these are just traditional machine learning. Let's say you're gonna have a, Tree-based model that, you know, use that data to create, you know, classification or prediction. Also Gen AI, right, using that data to either create an additional context for Gen AI, also just to make sure that you can prevent almost hallucinations saying that, you know, do not go beyond that. Here is your data and try to extract insights out of that. So all that processing is done over there. Then you gotta push this tool.

AI assessment note: “Yeah, ZDP is built on Google, BigQuery right now.”

Q Uh, this is your second company, right? You did Paribus before. Maybe talk about that prior venture, and then, um, how did you get to start Ramp? Why Ramp, and how did it all come about?

A Yeah, yeah. Um, so Paribus was about a decade ago. Um, if it was a 20, 24, we would have branded it as an AI agent. Um, but it was a weird savings app. So basically it was an app that lived in your Gmail or Yahoo. And the premise was, let's say you bought something at Amazon, Best Buy, Macy's, whatever. Um, let's say you bought a TV for a thousand dollars. The next week it goes on sale for 900. Every store would guarantee that you could get the difference back if you asked. We built an app that asked for you. Um, it would detect receipts in your inbox, go and scrape and track the prices, um, ingest the policies, and if you were eligible for money back, it generated an email as you, um, sounded like you wrote to the store, uh, chatted with their chatbots, and you would, as a user, wake up the next day to a hundred dollars or whatever back, and we charged a percentage, and, um, really fun, sort of insane, weird business, but we, we, we launched it in 2015. Within a year, uh, we had about a million customers, and Um, originally we wanted to partner with Capital One, and, ah, they said, you know, we'd actually like to, to buy the company, and so that's how we ended up there. Um, and in many ways, it was kind of the precursor, um, that led to what RAMP is today. Um, we learned a lot about turning data, ah, into savings, um, to go trigger some action based off of it. Um, We ended up …

AI assessment note: “in many ways, it was kind of the precursor, um, that led to what RAMP is today.”

Q Yeah. Do, do you, do you use like any of the systems, uh, for the book, for your articles, like as a, as a journalist and editor?

A Um, so, uh, yeah, I figured that would be a question people would ask me. Um, I did not, I really didn't use it for the book at all. Um, I didn't use it to do any writing. Um, and in part cause I just find like for writing a book length thing in the, in my own style that would sound like me and I felt like would be have the quality that I demanded. Um, I didn't find the systems that useful actually. I thought they were really good at like, they're really good at crafting a business letter. They're really good at Writing a quick email response. Um, I at least struggled to get them to write something that sounded like me. Um, so that was the problem. Maybe I, I don't know, maybe I, it's interesting cause I've, I've talked to Reid Hoffman a number of times and he seems to have had more success in getting, uh, fine tuning his LLM to, to actually imitate his style. I've, I found it actually kind of difficult to take, uh, a current system and get it to, uh, match my style properly. Um, so I didn't really use it for the book. I, I occasionally use it in my work for phrase finding. It's very good if you're, Sort of stuck for a phrase, and you know, if you're stuck for a word, or if you're overusing a word, you can always go to thesaurus, and that's very quick on, you know, on Google, you just go to thesaurus and look it up. But if it's not a single word, if it's actually like a kind of…

AI assessment note: “I did not, I really didn't use it for the book at all.”

Q So what about, uh, self-hosted, uh, deployments? Are you, uh, strategically deciding to not support those? Are you supporting those in some way? What's the, What's this thinking?

A So we do work with a handful of companies that are using ClickHouse in a self-managed way, whether it's on-prem or they're deploying it in a public cloud provider, but they're not using our hosted service. And an example of that would be Netflix, for example, that that's using ClickHouse, um, in AWS, but they manage that environment and we work with them. We provide technical support and there's a dozen other Companies that we also provide technical support around, but it's not our primary business model. Ultimately, we want to understand their use case so that we can improve upon the technology, make them successful. Ultimately, we'd love to migrate them to our cloud offering, whether it's the current multi-tenant serverless offering or in the future deployment model that we're going to launch later this year called bring your own cloud, where the data plane will sit behind the customer's VPC. We'll still manage the control plane. And so for data residency requirements, data locality, whatever its costs, security concerns, and we think that's going to open up a big market for us, especially in the enterprise.

AI assessment note: “we do work with a handful of companies that are using ClickHouse in a self-managed way”

Q And any stats you can share about the business, number of users?

A Yeah, yeah, yeah. The, the stats are, you know, 500,000 organizations use Airtable. Um, 50% of, of the, uh, Fortune 500 are paid customers. Um, and really the largest enterprises have been the, the biggest focus for us in terms of revenue growth and, and go to market, and even like the product roadmap, be able to scale up to support those customers. Um, you know, we, uh, if we were a public company, so we're not, um, but, uh, you know, we're in the hundreds of millions of revenue, and we would have ended Last year as a top decile grower amongst public SaaS companies. Um, and we also are forecasting a next 12 month revenue growth rate of, uh, also top decile, uh, growth.

AI assessment note: “500,000 organizations use Airtable. Um, 50% of, of the, uh, Fortune 500”

Q And in the same vein, uh, you know, Building data businesses, data infrastructure businesses, um, anything in terms of like the pace, the, you know, the do's, the don'ts, the positioning, like any kind of like pattern you've seen, uh, around successful data businesses or data infra businesses?

A So I think data is a much more ecosystem driven space than a lot of other spaces. If you're, uh, and what I mean by that is like, if you're a database, you need the upstream ETL vendors, you need the DBTs, uh, and the Zbicos in, in, in the center of the, and then you need the BI vendors. When Looker went to market, we initially partnered with Snowflake, and we were bringing Snowflake into deals, and that worked really well, and then Snowflake started to grow really fast, and Snowflake was bringing Looker into deals, and so, um, so why is this? Well, data is an ecosystem where there are lots of individual points, point products, but what a buyer wants is an end-to-end They, they want me to solve the problem. So that Looker or Snowflake combination was, I need a reporting stack. I need this report. And it needs to be this fast and be shipped on this day, and it needs to be in this format delivered to these people. Well, you need a database plus a BI. And so figuring out, like, what is that end to end stack that you're selling and becoming part of that, that stack is really critical. It really helps quite a lot. And the way that it starts is, It's typically AE to AE, so Looker AE plus a Snowflake AE get together, and they start selling two or three contracts, and then friends of that account executive say, hey, Matt, how'd you hit your number last quarter? And they say, well, I ha…

AI assessment note: “data is a much more ecosystem driven space than a lot of other spaces.”

Q So the database is to query the metadata, is that what you're saying? That's how it started, and everything expanded into.

A Yeah, everything in VAST is multi-protocol, and so, for example, you can write a file and then read an object. Everything is the same underneath the covers, and so the database is the same. You can write a parquet object and then query it using SQL natively off of our platform without needing any higher layers on top. Um, and so now people are using it not just to analyze their metadata, but also as a data warehouse. And in the same way that we broke those fundamental trade-offs, it grows to extremely large scale without compromising on performance, without compromising on consistency and ACID requirements. And so we find that in the database space, there are more trade-offs to be broken. Ah, row-based databases and column-based databases. That all stems from storage. It stems from hard drives needing to be sequential. If we can give different views into the same information, then you don't need that complexity. You can consolidate very similar to what we did, ah, in consolidating those tiers.

AI assessment note: “now people are using it not just to analyze their metadata, but also as a data warehouse”

Q All right, so data store, data base, and data engine, that's the compute part?

A Yes, that's where we bring it to life, and so both the store and the database are static in nature. You write to them, you read from them, you query from them. Our customers didn't like that. They don't like the fact that their application is written in that way. In fact, they want it to be, I'm going to put a plug for you now, data driven. Um, they want everything to be, ah, based on the information as it flows in. That genome file came into the system. It should trigger that inference function that runs on it. It should trigger an incremental training, ah, job that runs on it. That should run on a low-end GPU. This one should run on a high-end GPU. Um, if we can build that language of triggers and functions where you can then Understand more and more about your information, and put that in the database, and that triggers more functions, and you have this recursive machine that's all data-driven, um, then we have the full stack, and that allows us to do things that couldn't be done before.

AI assessment note: “Yes, that's where we bring it to life”

Q Interesting, yeah. I guess the timing was perfect, right? Because you were, you caught the alternative data Uh, trend in hedge funds at the time. Okay. Very interesting. All right. Um, how do you sell today? Like you have a sort of a classic kind of, um, high touch sales organization kind of thing?

A We do. Uh, we are very close with our customers, and I would say the two things that we do well is R&D, building a product, and sales, selling the product. Uh, everything else we're not very good at. Um, We want a very intimate relationship, especially with the first several hundred customers, uh, which is where we are today. And, um, now through partnerships, we're able to get reach into everybody else. And so you mentioned HPE. HPE has, uh, tens of thousands of sellers out there, and they've decided to standardize on VAST as their file offering and, uh, soon to be other parts of their stack. And so they are Giving us reach, uh, beyond those most data intensive organizations. We're partnering with, um, AI clouds to get reach into these, uh, smaller AI startups before they get to a hundred petabytes. Um, and again, the analogy of the operating system, we want to, on the one hand, make it easy for applications. On the other hand, have everyone underneath us help us sell.

AI assessment note: “We do. Uh, we are very close with our customers”

Q Maybe, uh, take us, um, a little bit behind the scenes in terms of the core infrastructures. We have this global audience, uh, that interacts in real time and creates three D worlds in, in real time, um, from a, uh, yeah, core infrastructure perspective to scale up and down. How does that work?

A So it's an incredibly complicated in-house technical stack in order to present a really robust and simple interface to the user. So we have our core data centers. Um, we have about a 100,000 servers that we maintain with our own infrastructure, um, running our own software stack on those. And that that's the heavy lifting backend. That's how we persist all of the data. Um, you have an avatar in the virtual world. It's how we keep track of what you're wearing, what the social graph is, um, all of the items in the world, all the three D worlds. We have edge data centers around the world. I think we're up to 17 of those, and so wherever you are, there's geographically located near you an edge data center so that we can have low latency, so that the experiences can be really responsive, because there's not a lot of telecommunications delay. And then we, of course, have the clients, which run on every device, so you could have a game console, you could have a VR headset, a lot of our users are on mobile devices, especially phones and tablets, Laptop, a desktop. And so we present whatever device you're on, wherever you are, we present a window into this three D world where you can interact with real people.

AI assessment note: “we have our core data centers. Um, we have about a 100,000 servers”

Q Full production. Okay. Wonderful. So, um, so what, what does it do? You alluded to this a second ago, but what does it do?

A So AI code assist, um, It's primarily targeted at helping people who are new to programming on the platform to quickly become better programmers. So for this specific feature, sort of setting aside the long-term vision, this specific feature is not enabling no-code creation of sophisticated new algorithmic development. That's not our intent. It's to say a human being is going to be the software developer, but we want to remove a lot of The sort of the learning curve for them, and we want to automate a lot of the tasks that might be more boilerplate. So with AI Code Assist, you're sitting in your text editor for writing software code inside of our studio tool. You start to write what you want a piece of code to do, or you start to write the actual code itself, and it will recognize and sort of like an email when it auto suggests, you know, do you want to say yes, absolutely, or no, I can't make it tonight. It auto suggests, but instead of auto suggesting a sentence, it auto suggests 10 or 15 lines of code that perfectly fit into your program. So it might be the equivalent of you're writing a leaderboard for a game. Maybe you're making, you know, Virtual field hockey. And it will, if you're starting to look through the players and assign scores, numbers, and it will recognize what you're doing and say, hey, here's an example of how do you iterate throughout the player. Here's how…

AI assessment note: “instead of auto suggesting a sentence, it auto suggests 10 or 15 lines of code”

Q So still following the chronology, um, AI image generator, uh, was, was the, the next one. What does that do?

A Yeah. So AI image generator, um, so you've probably seen tools, um, You know, for making sort of clip art stock photos where you, you type in a description, like, you know, cathedral with Socrates or something, and it, and it produces some image, right? Uh, and they're a lot of fun to, to play with, and there's a lot of services like this. Um, so what we were targeting instead was, because our environments are three D, um, one of the building blocks of making a three D scene is that you separately define the shape of objects, the geometry, usually by a polygon map. And at this point, I think everybody has seen the sort of like, you know, white outline, wireframe, kind of behind the scenes stuff. So you just, you define the geometry. Then you essentially paint the geometry with materials, and that a lot of the detail is not actually in the shape. It's just sort of painted onto the surface as a, you know, texture, so to speak. And there's a couple levels of that. One is, it's, it's literally the color that each pixel will be when the object appears on the screen. The more sophisticated level, which is what most of Roblox's technology is based on, is called, uh, physically based materials. And the idea is instead of painting the colors of objects, you're painting the chemical and physical properties of the material, how it's going to reflect light. Is it smooth and shiny? Is it, i…

AI assessment note: “painting the chemical and physical properties of the material, how it's going to reflect light”

Q You've been building this location data asset for well over a decade now. Can you give us a sense for the breadth and the depth of the location?

A Yeah, I mean, I think one of the great things about, um, you know, uh, the legacy that we get to build off of in the social app context is we were crowdsourcing effectively a map of the world. Um, and so we operate in a 190 countries globally, You know, two hundred million global POI, or places of interest, including a level of specificity of knowing the difference between Foursquare's headquarters and being in Fat Denny's cafeteria, and having that telemetry of what's in the world, um, and then building mechanisms that keep that fresh. The world's obviously super dynamic in terms of how places change over time, um, and so the starting foundation of what we do is really understanding Places in the world at scale on a global basis. And then it's attaching in real time how devices move in relation to those places. So what does that mean for all sorts of pattern, uh, development for understanding foot traffic? If you're a retailer, uh, you know, you can think about applications in places like logistics or real estate understanding neighborhood patterns or, or so forth. And so that's the, you know, sort of nature of creating this very dynamic asset, which is constantly changing. And building the mechanisms to be able to have what we think is unique other than basically Google in the world, which is human confirmation. We have elements of humans telling the machines that we're right…

AI assessment note: “we operate in a 190 countries globally, You know, two hundred million global POI”

Q And you can create video using Runway with eight different kind of inputs?

A Yeah, that's critical for us. We always think about Um, we're thinking about, like, who's using this, and for what? And if you're creative, if you want to tell a story, if there's something you want to say, You want to have full control. If you don't have control over the way you're using something, then it won't probably like matter, uh, because it will be just like a system creating things for you without actually you being the one with, with the agency. Um, and so we have like nine different ways of controlling those models and not all of them will last. Some will change. We'll come up with new ones. And so we're experimenting with these new medium, but there is, for example, um, motion brush. So motion brush is this idea that We train a model and a way of manipulating video just by, this is inspiration that comes a lot from how filmmakers and art directors actually give reference on, like, films. When you're, when you have a photogram, you sometimes just take a pen and, like, draw on top of it and, like, define how you want movement to happen. And so we took that inspiration, like, you have a video, an image, you can basically draw on top, define how things you want to move, and then, like, the model will actually move them. And that, and so you have, that's, that's motion brush. Um, and you have different, like, like eight different pens you can have for that. Another one …

AI assessment note: “we have like nine different ways of controlling those models”

Q Okay. So, uh, let's talk about what Ezra is today in terms of products. Uh, so we have different kinds of scans. Do you want to explain what they are?

A Yes. So the way Ezra works is you go on Ezra.com, And you book a scan. We have three types of scans. We have a cancer focused scan, which is a 30 minute, uh, fastest, uh, full body MRI in the world. 30 minute scan focused on finding cancer. We then have a 60 minute scan that is a cancer scan plus musculoskeletal analysis, so it does hips and spine, so it will tell you information about, that's not If you're pertaining to cancer, that might be useful for your health. And then we have something called Ezra Full Body Plus, which is a full body MRI, plus a low dose chest CT for lung cancer screening and heart disease assessment. I can kind of dive into that. And so we've designed these different scans in order to cater to different price points. So the 30 minute scan is 1350, the 60 minute scan is 2000 dollars, and then the, um, full body plus, the most kind of comprehensive one is two and a half thousand dollars. Um, so you go on Ezra.com, you book one of these scans, You visit one of our partner facilities, and we don't own and operate facilities, we partner with existing facilities, buy MRI scanning time from them, and run our own protocols, AIs, software on their magnets. You get the scan there, and then three days later you receive a report. That's not just the radiology report, it's like a translation of what the radiology report, which we also do using AI, and we'll dive int…

AI assessment note: “We have three types of scans. We have a cancer focused scan”

Q And speaking of reports, I guess the last part, like the last AI was what does it do and how does it work?

A Yeah, so the reporter AI has actually had a significant improvement in efficiency for us internally at Ezra. The way the Ezra reports work is we don't want to drop a radiology report on you that has like all these technical terms and you don't understand half of it. We want to explain every single finding in detail and tell you what you should do about it. We used to do this manually. Our doctors, we have a team of primary care physicians internally who used to take the radiology report, spend 90 minutes per report to generate an Ezra report, which is on average about seven pages, to describe to the member what each finding means. We built an AI that is able to do exactly that automatically, and our medical providers went from spending 90 minutes to generate these reports to five minutes to just reviewing them. So it's had a kind of profound impact on our business. We, we used to spend, people on average would get their reports within five to seven days, uh, from their scan. We now deliver it consistently in three to four days. So it's kind of like, You know, an incredible efficiency for us, which has led to a significantly better experience for, for, for, uh, for customers.

AI assessment note: “We built an AI that is able to do exactly that automatically”

Q And from a fundraising perspective, uh, What has been the, sort of the journey and the learnings? It's, you know, it's, it's a notoriously harder area to, to raise what has been your experience.

A I mean, clearly you've had success, but we've been fortunate to be able to raise. So we, we had a great seed round with four million dollars. We had a great, um, series eight led by Rick, um, uh, sixteen million. And then now we've, we've raised another great round. Um, it is certainly harder to raise in healthcare than it is in Other areas probably because, uh, a lot of investors just don't understand healthcare. And so they get a little bit kind of you, you pitch them and you know, in the meeting that one, they don't really understand it. And two, they don't really want to get into the, um, kind of rhythm of having to get FDA clearances and so on. They kind of want to steer away from that. Uh, and then, so you have a lot of investors that are not going to engage because they don't want to go into the space. You have a lot of kind of life sciences, kind of biotech investors that are not going to engage if you're a more AI software focused company because they do a different type of thing. You know, they want pharma and drugs. They don't want kind of software and AI because that's what they know better. And so it kind of narrows the pool to, uh, Not a lot of them, and then within that pool, there are not a lot who are good. And so, as a founder, you kind of need to play the numbers game and meet with everybody, and, um, hope that in that process, you find someone that likes you…

AI assessment note: “for the Ezra raise, I had a hundred meetings. I pitched a hundred VCs”

Q And so what did the beginning look like? So you have this, uh, inspiration, this deep desire to pursue a mission. What did you do next?

A Yeah. So I was actually still at TEEDS and was kind of struggling with this, um, uh, Problem, both personally and my family and, and just thinking about it. And I, I started reading research papers and, um, as one does when, as one does in general. And actually the, the, the funny story is the, the idea for Ezra came on my honeymoon. I was reading research papers on my honeymoon as you do. And I will never forget a moment. I was reading a paper that was comparing MRI with low dose chest, low dose CT ultrasound and other imaging modalities, PET. In terms of their sensitivity and specificity, which is kind of a measure for accuracy in medical imaging. Um, uh, and this paper was concluding overwhelmingly that MRI is the best imaging modality. So I turned to my wife. We were on the sunbeds on the beach. I was like, hey, if you could do a full body MRI in an hour.

AI assessment note: “I started reading research papers and, um, as one does”

Q At what point did you feel that the product was advanced enough to have People, because obviously the, the beauty of this is that it saves lives. The, the, the risk of this is that it's, it's, you know, mission critical. So like, how do you think about the right moment?

A Yeah. So we took, it took a while. It took two years from having the idea to actually having a product in market. Um, and there were two things that we did that I think were with, with hindsight, like really smart things. The first thing we did is we built a large advisory board. We brought on board 22 scientific advisors Before we even launched, and these were heads of body MRI at Memorial Sloan Kettering Cancer Center, uh, Siddhartha Mukherjee, a Pulitzer Prize winner oncologist at Columbia, the chair of oncology at Columbia, um, some other imaging experts, and we built the AIs and the scanning protocols together with these, um, individuals. The second thing we did is we ran a lot of tests, validation tests, prior to launching Ezra with people that we knew had issues of potential cancer and so on that we scanned in order to ensure that we could find the lesions. And we did, and then we launched the product, and it was an immediate success. Both our prostate scan, which is original prostate MRI, as well as our full body, um, our first prostate MRI that we ever did, we found prostate cancer. And, um, there was a gentleman who actually was from Germany. He happened to be in New York. He saw this thing, he got a scan, and we found prostate cancer for him. And he sent me a heartfelt email, like a couple of months later, saying that we likely saved his life. And this is the first e…

AI assessment note: “It took two years from having the idea to actually having a product in market.”

Q Okay. Uh, lots of, um, really interesting things in, in there. Uh, so in no particular order on the price side, is that, um, covered by insurance in the process of getting covered by insurance or FSAs or?

A So not yet covered by insurance. We're working on it. Uh, we do do a number of things to make it more affordable. So you can use HSA FSA dollars. We have a firm so you can pay monthly. And the thesis for Ezra, actually from day one, it's kind of been a constant, has been, uh, build a scan that's high end, our 2000 dollar scan, use the success of that to build a scan that's kind of more mid-market, that's our 13 50 scan, our 30 minute full body, use the success of that to build a 500 dollar scan that more people can afford and that you can then obtain and pay a reimbursement for. And we're currently at the 13 50 scan, and within two to three years, With a lot more AI, uh, we think we're gonna get to a 500 dollar 10 minute full body MRI.

AI assessment note: “So not yet covered by insurance. We're working on it.”

Q Yeah. You know, your, your point about hallucination triggers the. Fairly obvious thought and question around, um, false positives and anywhere this can go wrong. How do you think about that?

A Yeah, so, um, there's the false positive in MRI, which is something that's kind of important to, to address, and then there's this false positive in AI related things. I'll, I'll, I'll tackle both. So, MRI is an incredible modality because it's highly sensitive. Sensitivity is the measure of how good is a test at finding the disease. A test that has a hundred percent sensitivity will just catch everything. Specificity is when you find something, is it the disease you are looking for? Is it specific to that particular disease? In our case, if we find something, is it cancer, or is it something else? And so, uh, MRI is highly, highly sensitive. Uh, 96, 97, 98% sensitivity. It doesn't miss anything. Specificity is probably around 80% to 90%, depending on the organ. And so, what we've developed internally to minimize those false positive rates caused by the 80 to 90% specificity Is a number of things that allow us to determine what we should follow up on. So for example, uh, every single Ezra finding gets ranked on a score of one to five. A, uh, that gets done by our medical doctors, but it also gets done by the AI. So kind of every report, the report translation that I was talking about, every single finding not only is a kind of an explanation of what it means, but it also has a score in the background. And that score determines whether we tell you to follow up on the thing or no…

AI assessment note: “there's the false positive in MRI... and then there's this false positive in AI”

Q And by the way, uh, what is the latest number in terms of resolution rate? I think I've read somewhere. 41%, uh, of queries. Is that higher now?

A It's a bit higher. It gets higher every, every few weeks. We, so we just shipped Finn earlier this week in 40 more languages and that will kind of give it a higher involvement rate and probably a higher success rate too, because it was probably, you know, speaking bad, bad language, like, you know, bad English or whatever in previous cases or whatever. Uh, so that number is creeping up. I don't have the current one, but like forty-ish is, is roughly where we're at. I will say like that, That kind of hides, like on average, I would be confident any business who turns on fin will get at least like 25, 30. We have a lot of people getting 70 and 80. Uh, it kind of depends on the simplicity of the support function that you're staffing. So if like you have some sort of Pareto style, 80, 20 for your support queries, as an example, like utility provider deals with like open an account, close an account, change an address, register a meter reading, whatever, uh, those four or five questions often account for 80, 90% of the entire inbound. In those cases, FIN delivers exceptional results, as you guessed, because you can target, you can deliver so much value by just getting really good at four or five things. Um, so sort of like it's, the 40 thing is like, is definitely like, you know, it's very, very real. There are just, there's a lot of cases where if you have a simple, simple enough s…

AI assessment note: “It's a bit higher. It gets higher every, every few weeks.”

Q And for Finn, you charge, uh, based on resolutions. What is that concept? Uh, and is that something that customers understand or need to be educated about in this brave new world?

A Um, there's a tiny bit of education. Like we actually consider resolution the same way they do. Uh, but you just have to explain a little bit. So, so let's say a thousand conversations come into a business. Uh, let's say Finn only touches 700 of them because it looks at 300 and goes, I don't know what to do with that. Uh, because maybe it hasn't read the right docs or hasn't been fed the right information, or maybe they're gobbledygook or spam or whatever. It doesn't really matter. It's 300 of those. It's not touching. That's not relevant. So the first thing we would quote is our involvement rate, which would be 70% in this case. 70% of the conversations Finn jumps into. That's not what we prize for. We prize for when Finn has given an answer and the customer has either explicitly said they, has either just closed the messenger and gone on and done the thing they wanted to do, or has explicitly said that answered my question. The only thing we don't charge, the only time we won't charge here is if the customer pushes back and says, that's not right. This is wrong. And we hand over We're human. That's what we don't charge. Uh, we basically charge when we gave the customer an answer that they saw and they didn't have any follow-up questions, which is exactly how CS reps are measured as well. And that like, no one goes chasing people being like, are you sure? Are you sure? Are you…

AI assessment note: “Um, there's a tiny bit of education. Like we actually consider resolution the same way”

Q Um, and those are your own models, uh, or like, how do you go about building those?

A No, it turns out that the hard problem here is actually, I mean, there's enough SQL written in the world. There's enough. Even dbt code written in the world where just your standard foundation model will, will do this stuff pretty well, which will be interesting for us to figure out, like, can we over time scale this stuff with lower cost, whatever, like the right now, you know, you plug in GPT, 3.5 and it's not as fast as you want and it costs more than you want. Um, but like something is as stupidly simple as like, Uh, insert formula and you describe the English text to what you want the formula to do and, uh, having it write the regex without you having to go to stack overflow and say like, remind me how to do email splitting regex. Uh, it's just such a big performance accelerator. Um, and so I think that, I think analytics engineering is going to Change meaningfully over the next two years. And beyond that, I have no idea what anything is going to look like. So I don't try to predict beyond that.

AI assessment note: “No, it turns out that... just your standard foundation model will, will do this stuff”

Q And, um, so how did you finance it? You know, bootstrapping is always a fascinating topic to me, which is maybe weird because I'm a VC, but, uh, like, how did you, was that services, or, like, how did that, how did you manage to do it?

A Uh, so it's obviously a very constrained environment, right? Um, the good news is it didn't take us long to build the first product. And I think that's a lesson, right, that has been also taught by many people, but I think it really, we really felt it, like the first product we built in three weeks. And I still remember I was there, I was contributing a little bit, and then Martin, who is our, is our key engineer, he's our CTO, was writing most of the code, and then Basti was bringing us beer and pizza, and we were literally in the office for three weeks, and we had our first prototype. It was not very sophisticated, but kind of did the job. You could do demos, and then we went out, and we tried to meet with customers. We, like, used every One of our friends who did an internship in some company to get like an intro and get a meeting, and then in the first year we found, I think, five or six customers that were sort of willing to, I don't know, pay us 20 grand for a pilot or, you know, 30 grand for a POC, like some, you know, not like huge projects, but like real, you know, money. So, um, so, so we had some money to hire our first employee and, you know, Buy out our first booth at a conference. I mean, the first conferences we went to, we couldn't even afford the booth. We just went there and talked to people, right? Bit awkward, like, hi, I'm the guy with the business card, an…

AI assessment note: “we sort of worked our way there through funding customers”

Q was great. Uh, the, the open source, uh, AI world right now seems to be just completely exploding. It's super exciting. Is that, is that, uh, like, for, for you guys who are, like, deeply into the space, is that, is that the same impression as, like, the, The rest of us, like what's your overall sense of the, of the health and vibrancy of the open source AI ecosystem?

A Yeah, it's, uh, all time high as I would say right now, like we've got a ton of VC money flowing in the, you know, companies that are shipping things open source. So like, I love that that's happening. I love that, you know, I think on a podcast right after GPT for all came out, I think it was the weights and biases podcast with Lucas. I said something like the biggest challenge for open source is going to be getting like the monetary resources to do like a seven billion model or these bigger models. And so it's amazing to see companies like Mistral actually going and doing what they say they're going to do and releasing these amazing models. And Meta as well with the release of Llama I think has played a big part here. And I think also one of the reasons it's kind of popping off right now is we really are at this kind of new frontier in terms of the discovery of what these things can do. And so it's totally possible that some Random person, you know, in the middle of nowhere that's just like somewhat interested in this stuff, you know, spends enough time poking at it. There's so much new stuff to find that they can discover things that are really amazing. And so you see some of these really incredible techniques coming out of, you know, not maybe what you might call like the Royal science or like the, the universities, but some random person will be like, oh, hey, like I did t…

AI assessment note: “Yeah, it's, uh, all time high as I would say right now”

Q Any, any, any lessons learned there from a go-to-market perspective? Is that, is that just, um, you know, generally comparable as selling developer tools, meaning you need to be authentic and you need to be responsive, or what have you learned?

A Yeah, so I think one of the most interesting things I've learned was be open to your, your thesis about core audience shifting. So when I originally built, you know, and envisioned Atlas, Um, and sort of brought it to Andre and the team to, to help me realize it. I very much thought that the, you know, ICP ideal customer profile would be a machine learning engineer. And, you know, a lot of our early users were machine learning engineers. We still have a lot of them using the system, but something that struck me as very interesting is Atlas started to get picked up by a lot more, um, kind of like, I don't want to say less technical teams, but an archetype that's more of a business analyst, um, or someone doing business intelligence where, Um, we see, you know, consulting companies that get these data sets from their, from their customers, and maybe they have domain expertise, but they can't code. And so never before have they been able to interact with this data, uh, with this level of velocity and this level of granularity. And so we've seen, you know, a much wider adoption in terms of how technical the, uh, the actual users of the systems are than I originally anticipated, which has been really, really interesting to see.

AI assessment note: “be open to your, your thesis about core audience shifting”