Starting & Scaling Micro1 To $200M+ In Revenue With Ali Ansari | EM Group Chat #207

Welcome back to Everything Marketplaces, where we talk with founders and leaders on some of today's top marketplaces. So this is episode 207, which is really a group chat I just had with Ali Ansari, who's the founder and CEO of MicroOne. MicroOne is a platform that helps AI labs and companies find, hire, and manage human experts for data labeling and training. They've been noted as one of the fastest growing startups recently and scaled from a $4 million run rate to over $200 million in the past 12 months. So it was really great chat with Ali. We went back to the very beginning and how MicroOne initially started as an AI recruiting tool. Ali shared how it evolved into a marketplace, walked us through what the platform experience is like for both AI labs and experts. We got to learn about their incredible growth, how they're leveraging AI to operate and scale so efficiently. We discussed topics like betting and matching, the future work with AI. We also had a great group Q&A. I really enjoyed this conversation and I think you're going to find it a great watch to the end. I'm really excited to have you join us here today for the group chat. I want to start off by saying huge thanks for taking the time to do so in advance. Of course, we have a lot that we're going to dive into with MicroOne and your journey, but I thought it'd be great if you could start off by briefly sharing a background here for those that might not know you.

My name is Ali. I'm the founder and CEO of MicroOne. Studied computer science and math at Berkeley, and it's where I actually started MicroOne. It started as something that's quite different than what it is now, which we'll get into. But yeah, excited to be chatting.

Thanks for sharing with us on that. And I guess for a little bit of context, can you give us a brief overview of MicroOne today? And before we dive into things.

Absolutely. So MicroOne is the AI platform for human intelligence. So what we do is we vet highly skilled people, mainly PhDs, professors, and industry experts across many domains, but mainly medical, finance, and legal. And we help train AI models. So you can sort of think of our product as a data engine product that labs use to consume kind of post-training data from. And there's 3 core pillars to it. First is our AI recruiter. Which is an agent that essentially sources and vets different experts and adds them to our talent pool/marketplace. And then a second is our data platform, which is where the data is actually created. We have a bunch of AI-based kind of QC mechanisms built in to assure quality. And then lastly, we have a product called Merit that is our talent performance management product that essentially quantifies expert quality on, on different data pipelines for like data quality being higher, for velocity being high and so forth. And all these together allow us to kind of help train AI models at frontier labs.

It's going to be a really helpful context for us here. And I think a question that you probably get quite a bit, and a few here are probably asking is, why don't the labs kind of do this themselves? Why is outsourcing it to a platform the best solution?

There's a few reasons, and most labs have tried to do it in-house, but they've decided basically not to. And there's 2 kind of core reasons. One is The technology, which is, you know, the sort of sourcing and vetting, the data platform, talent performance management, all the rest is obviously difficult technology to build, but it's also a distraction for foundational models that are building LLMs. And the second thing is if you're hiring a bunch of experts in-house, you're, you know, the volume that labs hire is in thousands of experts. And if you're hiring thousands of experts in-house, you're really kind of changing the composition of your internal team from like researchers to just a bunch of different domain experts. And you really kind of start to become not a research lab anymore. So there's, yeah, these are the 2 core reasons, but there's a bunch more. And really the, I would say the 3rd one is like the labs heavily have to focus on operations and like data operations if they build some such thing in-house, which—

Yeah.

Which we sort of handle.

And, you know, I'm sure it's not like a, you know, easy to kind of break into, right? So I did want to go back to the very beginning. You know, what was the founding story for starting and what were the first steps that you actually took to start at the time?

Yeah, so when I was at Berkeley, I was developing a software development agency and it was a very service-oriented, you know, we just built websites and apps for other companies. Nothing super special. But one of the main kind of limiting factors at this agency was Every time a project would come in, I had to assign engineers to it and I had to vet and interview a bunch of new engineers to be able to assign to the new projects. And so a lot of my days was spent essentially on interviews. And so I was thinking of ways to, you know, solve this kind of bottleneck that we had and built this internal tool, which was an AI screener that helped me, you know, vet engineers. And then pretty quickly, this AI screener that I built as an internal kind of use case became MicroOne. And I ended up kind of selling the tool as a software. Like I started licensing the tool to other companies to use for their recruitment. And then we also started to build this marketplace of this kind of like next generation marketplace of pre-vetted talent that has already gone through these AI interviews. And mainly the talent was focused on engineers, but then we We kind of quickly generalized to like product people in general, not just engineers, and startups would hire from that marketplace. So we had this 2 sides to the business model. One was the AI recruiter software, you know, people subscribe to it. And then we had the marketplace that startups would hire from. And then about a year or 2 into the business, we realized that the human data world has kind of a new bottleneck, which is the recruitment of experts. And really requires a product that has built already a talent pool that also has the ability to source and vet very quickly any type of talent. So long story short, we decided to basically go all in on that application of what we had built. And now we, you know, we were sort of building data infrastructure for AGI.

And, you know, when you did kind of transition into them and into the marketplace model a little bit more, you know, like what were some of those like initial challenges that you faced?

You know, the challenge with any marketplace is Kind of 2 core things that we were having. One is you have to have supply and demand, you know, at the same time. And when you don't have one, the other is hard to get. And we were facing a little bit of that early on where we actually didn't have enough demand for like engineer hiring because, you know, we had to kind of build up the client base. And so the engineers that would go on the talent pool, you know, oftentimes would be not happy because the duration that it took for them to be placed was kind of longer than expected. So those are, you know, those are challenges that we had to kind of solve. And then I think the other thing was we, the first versions of our AI interviewer were not, you know, they were not the most accurate. So we had to actually have human interviews happen afterwards after every single AI interview, which was obviously a very large bottleneck. Like there was still a human bottleneck to build this marketplace. So we had to, we had to really get the product to be very good in terms of vetting talent to remove the human interviews. And then, and then that really unlocked the supply side even more. So, so yeah, those are kind of the 2 main things.

Certainly the cold start problem is something we all, we all face. And then, yeah, I think for more kind of like traditional, I would say, you know, recruiting and hiring marketplaces, I think one of the challenges, like as you know, you know firsthand, is not only that kind of sourcing aspect, but then like the vetting and kind of matching. And I think it's really interesting here, you know, you're approaching 2026 with AI, right? Like building an AI-native marketplace. So could you kind of like walk us through that, how you think about that, like that sourcing and then that vetting and getting to a match?

We look at the talent pool that we have, talent pool slash marketplace that we have is very kind of AI native. And what that means is we have an AI recruiter that essentially does the sourcing and vetting for the marketplace. And the sourcing part is a bit more simple. It's 2 main things. One is programmatic job postings on different platforms on where the jobs are going to kind of perform best. And then the other is, and this part is actually not Agentic, which is we have a referral program and experts that are already active, they refer others. And there's an amazing network effect here, much more than we actually, much more than we thought. We initially, we were having, if you look back like 3, 4 months ago, the referral program was about 30% of all of our hires. Now it's about 50% and it's growing in percentage really fast. And I think, you know, in the long run, it's gonna be basically close to 100% of our hires will come from referrals, which is, which means the sourcing mechanism is just completely kind of built in, built in-house with the referral program. So that's the sourcing. Now the vetting is the part that we really emphasize and we really kind of build a lot around. And there's 2 components to the vetting. The first is our conversational interviews that happen. which is basically the expert conversates with our agents about, you know, skills that they should have for any given job. So if it's an engineer, they're gonna be tested on, let's say, React, Node, Python, whatever it is. And each skill is about 5 to 7 minute conversation, just like a human interviewer would interview an engineer. And it sort of starts with a question and based on how the expert is answering it, it goes into different directions and kind of digs deeper and has a really a natural conversational flow. And then after the conversational part, there's a, there's an exercise that happens, which is, you can think of it as like a take-home almost exercise, but it happens live. And we, for programmers, it's of course a coding exercise, but for every other domain, this doesn't really exist. So what we've built is a model that can generate exercises for any domain. So if you have a lawyer that's being tested, they are doing some sort of like redlining and there's like a contract interface and they're building rubrics to actually allow for the redlining to be, you know, to be good redlining. And then if you have like a doctor, maybe a radiologist, maybe they're looking at some image and they're analyzing it. So basically we've built this model that can generate exercises on the fly that's completely customized to exactly what the job needs. And then this, this allows us to really deeply vet any given candidate. And then once they go through the whole process, which is the conversation report and the exercise, we generate a report on how they did. And there's basically a, on each skill, there's a rating. And then, and then there we have recruiters that look at those and kind of determine the final fate of the candidates.

Yeah.

And you mentioned network effects earlier, which is like a fun topic that we're always discussing here. So yeah, I guess like on that note, you know, when did this flywheel really start to kind of kick in here? And like, what do network effects look like for Micro One?

I think it was the first maybe week or two of launch was already kind of looking like 10 to 15% of our hires on a kind of a projected basis. And then within a couple of months, it was already 30% of all hires were coming from referrals and now it's, you know, above 50%. So it's a product that we actually should have launched a lot earlier. We, I thought that it would help with sourcing, but I thought it would just be kind of a small part of all our sourcing strategies. but it ends up looking like it's going to be the main and maybe even the only sourcing strategy that we need. So I, you know, I think part of it is you have to, you have to have a high NPS score. You know, obviously the experts need to have a high probability of referring their friends, but it, you know, if that's true and if there's a lot of kind of happy folks on the marketplace, then I think— referral program is, will likely be a pleasant surprise.

Certainly. And I actually had a note for a talking point that I want to discuss a little bit later, but it was around that happiness and how you're actually measuring that on the supplier kind of expert side.

One of the things that we emphasize heavily is just taking a very human-first approach. There's, I think there's something that's sort of undermined in this space a little bit. The people that are, the way that models improve is really just people. Like there's many, many people across many domains that are distilling their structured human judgment into these models. And then the models become more useful for the rest of humanity. So this is a very important job and we need to really treat it as such. So the approach that we take at MicroOne is putting the expert experience as like the number one thing for us. And we really build our products around that. So I'll give you, I'll use some examples on the recruitment phase. Once the experts finish the interview, there's a rating that they have to give, which is, you know, did you enjoy the experience? And, you know, what do you rate it? And if you have any comments, you know, leave the comments. And we've had this rating go from, I believe a while ago, from like 3.8, 3.9 to now a 4.5 out of 5. And this continues to rise pretty quickly. And, you know, we essentially take the feedback at every step of the funnel and really build our sprints around candidate experience. And the belief that we have is long-term, the company that wins this kind of human data space is the one that attracts the best talent and is really the one that makes the the best place to train AI models. And the labs will go simply to where the talent is, where the best experts are. And so, you know, if we're that company that has the most exceptional experts, that has experts that are kind of loyal to the brand, then naturally the labs will, you know, will have to sort of go to us mainly. So yeah, you know, those are some of the ways. And there's also this notion of, incentivizing our project leads based on the expert happiness index, which is as this measure kind of goes up, there's like the commission of experts or the commission of project leads actually are variable based on it and also go up. So there are these sort of mechanisms that we have built in the process, but we also have a— our model also looks at the probability that an expert will be satisfied on any given job and considers that in the overall AI match score before the candidate is actually matched with the job.

I'm sure it'll lead to a few questions once we get get to the group Q&A. But, you know, if we kind of jump over to like the job side, you know, what were those like initial like jobs that you were working on with, you know, like the AI labs? And then how did you think about like expanding and kind of adding new types of jobs?

The jobs are in a very wide set of domains. The way to think about kind of AI training in general is there's an expert that is trying to find something the model's not good at, which is a task or a prompt that the model fails at in their domain of expertise. And then of course the data structures and kind of pipelines vary, but for the most part, they're creating rubrics that define what a good answer would be. So model comes up with an answer, expert determines this answer is actually not good. And, you know, the expert essentially stomps the model and then they come up with the rubrics that define what is an actual golden response. And a rubric is essentially a set of, or rubrics are a set of binary statements that essentially determine what should be included in the answer. So if you have a, let's say a model needs to do some sort of redlining on a contract and like give a, as a response, give some sort of memo as well as the actual redlining of that contract, the rubrics would define, you know, the specific redlines that must have had happened. as well as what should be included in the memo. And this, you know, each data point often takes anywhere from like 10 to 20 hours to create. And this creates verifiable rewards for AI models to then to be evaluated, but then to also then train on that data. So this is the type of work. And what we started with is finance. We've had a pretty heavy focus on finance. And then now we have 100+ domains that we're working in active pipelines, but there's kind of 4 or 5 main domains that take a good portion of all the work that we do. And those are medical, finance, legal, coding. And then there's a lot of kind of voice experts, like voice actors and these sort of adjacent roles.

And I believe too, didn't you recently expand into robotics?

Yeah, we're also in robotics. There's 3 data verticals that we have broadly. First is what we call Realm, which is frontier evals and RL environments for foundational models. And then the second is what we call Cortex, which is contextual evals for enterprise agents, which we actually had a launch today about that. And then third is robotics, which is real-world robotics data for models that can act in the physical world.

Very cool. And I did want to spend some time talking about your growth, which is really incredible. And I know you've shared about it on social, I believe, going from $4 million to over $200 million in run rate. Can you walk us through, you know, how you've approached growth?

So yeah, we started 2025 with roughly $4 million run rate. We ended just under $150 million. And now in, you know, 2 months into 2026, we're actually quite above $200 million. We're now approaching $300 million. So growth is accelerating very fast. I think, you know, obviously we're, you know, we're trying to execute well, but we also have to give credit to the market. The market is There's an insane amount of demand for data. And if you can, you know, if you can recruit the right experts and have pipelines set up in a way where you can assure quality and just provide lots of high-quality kind of expert data, which obviously is not easy to do, the labs will buy it. And there's, you know, they're asking for more in every pipeline. And of course it's not just the labs now, it's enterprises that have labs within them, but it's also enterprises that don't even have labs. They also need evals for for agents they're building. So everyone needs data and everyone needs kind of expert level data to get agents that work in this context of actually useful work. So, you know, we, the approach we take is a very, is actually a very non-salesy approach to our go-to-market. We don't have that many customers, like we have like, you know, 10, 12 very important customers. And so we take the approach of building real trust and real relationships with them versus this sort of aggressive selling motion. So I think that's one. And then, and you know, that results in just natural demand that comes as we provide good quality. And then the way that we keep up with this demand is we, Try to productize our operations as much as we can. So instead of hiring more recruiters, we make sure our AI recruiter can just do more over time. And, you know, having— we have our project leads, essentially we have this key metric that is how many pipelines does each project lead own? And we make sure that this number is like quickly increasing, which means our product is doing a lot more of the work. So these are things that we sort of focus on to allow us to Actually be able to capture this demand.

That's really interesting to hear about. And yeah, I was going to ask one of my— one of the questions I wanted to discuss a little bit was how you're leveraging kind of AI as far as for like internal operations and efficiencies as you scale. Because I can only imagine things start to break with growth sometimes. But this is like unprecedented kind of growth.

We actually recently centralized our AI department to really help like across the org. Of course, the AI department mainly focuses on the 3 kind of main products that we're building and really maximizing kind of automations there. But there's a lot, there's also a lot of one-offs that certain teams need that actually help them a lot. And so centralizing the AI department was a move we made recently that kind of allows for this kind of more AI throughout the org. But there's, it's really the 3 products that we have that allow us to do this. The, of course, the AI recruiter, but then the data platform is another one, which is We, you know, it's where the data gets created and there's a lot of human review mechanisms built in, but we complement those with AI-based review mechanisms. So we have this tool called Rhea that essentially looks at data instruction guidelines and then looks at data points that are being submitted by experts and assures that data structures are actually being met and gives instant feedback to experts before it goes to the human review layer. So this is a product that, you know, this is a model that we built and that we kind of continue to iterate on that helps us increase the efficiency on the pipelines. And then, and on the Mirror platform, an example would be we, it's basically a talent performance management platform that showcases each expert and on how they're doing. And then project leads go in there and based on how experts are doing, they perform actions like, bonuses, you know, raises and potential replacements. And as we get this platform to be more and more sophisticated, it will automatically suggest these actions and project leads would go in there and just kind of confirm. And so these things are really important for us to kind of continue developing on and making pipelines more efficient.

Certainly. And then, you know, on the note, as far as like, you know, team building and hiring, you know, how do you kind of approach that? And then I would say, you know, kind of creating this like culture of, you know, moving so fast.

We try not to default to hiring. This is something that we've built in the culture pretty deeply, which is like, this is not exactly the right phrase to use, but like folks are almost like almost scared to ask me to hire. And I think this is actually a good thing because it means that our team is not actually defaulting to hiring Every time there's a problem and kind of trying to throw hires at it. So what we do is we try to keep the team lean and really hire when we absolutely need to. And when we do hire, we look for, you know, 2 main qualities. One is just pure agency of that person. Like, will they care and will they really put in the work that's required? And the second is, can we kind of see that there's a risk-taking appetite from the candidate that we're hiring. We tend to hire folks that are okay to take risk and are kind of bold even when they just start because the pipelines that are owned by our project leads and other types of people, they're really, they need to own the pipelines end to end and they need to make decisions without kind of asking for permission often. So these are 2 things that we look for when we hire. And we try to celebrate, you know, to kind of increase the risk-taking probability at the company. We try to celebrate mess-ups. I mean, of course not just like pure mess-ups, but like if there's a mess-up and there's a nice kind of comeback afterwards and there's, you know, more trust built with the client because of it, we really celebrate those and we make, you know, there's a case recently where there was a bit of a mess-up which then resulted in actually a lot more trust built after. We actually made this person that resulted in this mess up the team member of the month. So we do these sort of things that, you know, that kind of increase the overall risk-taking appetite for the company, which results in moving faster. And we look for those qualities when we're hiring.

And another topic that I wanted to briefly kind of chat about here before we get to the group Q&A is around fundraising. And most recently announced your, you know, $35 million Series A round and, you know, had some top investors on board. So, you know, what has the fundraising journey been like so far? And what are some of your, I guess, early fundraising learnings?

Across 3 rounds, we did pre-seed, seed, and then we did our A recently, like you said, raised total of about $40 million. Fundraising journey has not always been super smooth is the truth. I mean, especially the first 2 rounds that we did. The first round took, I mean, essentially like a year to close. Like there's a, it depends on like when you kind of call it kicking off, but like You know, when you start a company and once you've bootstrapped some traction, you start to get into fundraising mode pretty much right away and often takes a while to actually get it done. So, you know, first round took a long time, second round took a little bit less time, and then of course Series A was a bit more smooth. But each round has its own kind of difficulties. And I think the, I think one thing that it's probably not said often is like how difficult rounds are for like most founders. I think like once a round is announced, you know, you sort of put a bow on top of it, you make it a nice video and it all looks good and then it looks easy. But then I think no one can really see like what went into it. And so what happens is I think there's this wrong frame of reference that founders get because they see all these announcements. They're like, oh wow, it's like really easy for everyone else. Like, Why is it taking so long for me? I think that's actually the wrong way to look at it because those announcements, it wasn't easy to get to that point for pretty much anyone. And if your round is like not going well, it's okay. It's like the norm and at some point it will be done and then you will also put a bow on it and announce it and then it will also look easy. And then after some time you will likely forget how hard it actually was in retrospect. And that's part of the journey too, because now you're like feeling confident again for the next round because you sort of forgot how hard it just was. So these are things that I kind of remind myself as we go into new rounds. I'm like, you know, I think obviously growth is great and we're profitable and all these things, but still it's not going to be like a smooth sailing thing that we sort of think each time. It's going to be a helpful reminder.

I think a lot of times you just, you know, see the headlines constantly, right? Exactly.

Cool.

So we're going to get to the group Q&A, but right before we do, do you have any tips for founders building AI-native marketplaces in 2026?

The first version of our product was obviously this marketplace which companies could hire pre-vetted engineers from, but really the version that stuck well and resulted in this crazy growth was using this marketplace as part of our overall infrastructure that we've built for this one kind of use case. which is a bit more hands-on and a bit more of a white glove service that we productize in the backend. So I would say, you know, oftentimes there's one or two applications that are very clear that are working from the marketplace infrastructure you've built. I wouldn't be afraid to lean into those one or two applications and kind of make it a bit more niche versus, at least for some time versus like a, general marketplace buildup. That's something that did very well for us. But once we decided to go all in on the data use case, that's when really the growth started.

I think that's going to be super helpful. Cool. So we're going to get to the group Q&A. And I actually want to ask a question that was sent in advance to kind of help kick things off here. And it was actually around the future of work with AI. And so it's asking, you know, how do you think the future of work will start to kind of play out, you know, with AI and also, you know, potential paths of like AI training?

I think long term there's Likely a good percentage of the, probably anywhere from like 3 to 5% of all human labor will be spent on AI training. And there is, the way to think about this is if you look at the kind of current state of, let's use an example, lawyers, they, you know, they're spending time at their law firm creating unstructured data, essentially, like when they're working for their clients, they're creating some, they're, you know, creating some sort of redlining and creating some doc that they sent over email. And then that doc that was sent creates value in the economy, but then it's sort of lost after that and it's sort of very unstructured. Now the alternative is like working at a company like MicroOne and doing that redlining in a structured manner, which then models can improve from, which then help your job in the future. And currently the economy as a whole has essentially decided that lawyers should get paid about 20% more doing structured work versus unstructured. And, you know, if you look at our talent pool, we have about 1,000 active lawyers that are working on active, you know, on data pipelines and their pay is about 20% more what they make in their day job. So what this means for the future of work is that You have to assume that some, of course you cannot infer that because lawyers, for example, are more valuable as determined by how much they're getting paid creating structured data. You cannot assume that all lawyers should go and create structured data because then obviously the economy must continue functioning. And so, but I think you could assume that there's this equilibrium that we'll reach where some small percentage of all human labor across all domains We'll work on AI training, we'll work on structured human judgment creation for models to improve, to then kind of build agents. So that's the, you know, that I think is the main thing that we're trying to like capture, that market that we're trying to capture. And that's the main kind of change that we'll see in the future work.

I like that, you know, the way you kind of frame it as far as like, you know, structured versus unstructured data as well. So I think about it like that. Exactly. Hey, Neil, saw you raise your hand. Do you want to jump on?

Thanks, Ali. I wanted to come back to referrals partially because we have a somewhat shared supply side. Halo's an R&D partnering network, so we are connecting like big R&D orgs at enterprises like Pepsi or kind of traditional companies with researchers, typically PIs at universities or startups. So in that sense, like it's very similar, I think, of this of who you're reaching. We have not turned on a referral program yet. You mentioned NPS score and our internal debate is like adverse selection. So like what was the right NPS score of like, if you clear this hurdle, it'll probably work. And then did you experience any adverse selection or see a change that was not what you were hoping for in— Who was being referred?

It's a good question. The, in terms of NPS score, I don't think there's like a threshold. I mean, I think if your NPS score is super high, you're likely going to be very surprised with how important and how much of an impact the referral program will have. And if it's, you know, if it's a little bit lower, then it'll, I think it'll still have a good impact, but probably just a little bit less. So I wouldn't wait for, like a threshold necessary. But I think the adverse selection issue is one that we think about. There's also, there's an adjacent issue, which is just gaming the system and having folks that may actually refer in a kind of wrongful manner and kind of make money from it. So we've done the way that we try to fight against that is One, our vetting process, like, you know, we, everyone that goes through the, anyone that's referred go— goes through the same vetting. And if you pass our vetting, it's really hard to game the vetting. Like there's a proctoring model that looks at whether or not you're cheating. There's a conversational part, there's an exercise that takes about an hour to do and it really digs deep on your skills and it's very hard to game. Like you either have the skills or you don't and it doesn't care where you're from, who you are. Like, it's just based on the, the actual skill sets. And then the second part is that there's a mechanism we built, which is you cannot get paid out unless the expert actually works 10 hours. Once the expert puts in 10 hours, then it's considered hired and then you get paid out. And even if there's an expert that has gamed the vetting, if they join a pipeline, they will not survive being able to put 10 hours of work in if they're not actually good for that pipeline. And so these 2 things allow for us to not have this kind of biased selection as well as like reduce the chance of gaming the system essentially.

I'm gonna help ask a question that Iggy sent me in the chat. He said he's in a noisy space, but it was around more about the project leads. And so he was asking, you know, for a little bit more on like how they spend their time across, you know, clients with us, experts? And then, you know, what's been the biggest difference in optimizing how much volume they can handle?

Project leads are essentially what we call strategic project leads. They're essentially the ones that own the pipeline end to end. So on one side of the pipelines, there's some experts, there's a, you know, if it's a, let's say it's a medical pipeline, there's going to be maybe 200 doctors, maybe they're surgeons all around the world that are working on some you know, model improvement. In this case, the project lead will also likely be a doctor as well. We like to hire kind of expert project leads that understand the domain well. And, you know, some of the things they would do is they would set up instruction guidelines for the experts. They would set up our automated chatbot that supports experts called Lumi. They would help, you know, set up the data platform to actually be able to support the data taxonomy to start creating the data. You know, they sort of host office hours, kickoff calls, these sort of things. And then they, on the other side, they are interfacing with researchers and really providing updates on how the pipeline's going and what the quality of the data is and what the velocity is and so forth. And we, it's actually a really fun job because it's, you get to work with best-in-class domain experts. And then on the other side, you get to work with best-in-class researchers. So it's like a very happy middle ground to be in.

Certainly sounds fun. So, and then another question that was actually added in advance was around more so around like the moats and defensibility in the space. And so, you know, it's highly competitive. So how do you think about, you know, moats and defensibility for, you know, marketplaces in the AI data and training space?

The first and most important is actually around Around the marketplace that gets built. The network effect on the marketplace is the biggest moat. We, and very similar to that is us being able to build the best AI recruiter model. As we have more and more interviews done, we have a lot more candidates that go through our funnel. There's more active candidates that are working on the pipelines and then the active candidates on the pipelines give us ground truth performance data on the job. And that allows us to actually improve our AI vetting model because our AI vetting model is trying to predict the performance on the job. And so there's this like very nice data flywheel that gets built to improve our model. But also as we improve the model, we get more experts and the more experts start to refer more of their friends, which then again, there's higher kind of amount of top of funnel sourcing that happens. And this flywheel just keeps going. That there's, it's incredibly difficult to catch up to a flywheel like this that starts to scale. I mean, of course everyone on this call knows the sort of the Uber kind of network effect and the incredible moat that they have. Even with self-driving coming out, I think they have a distribution moat that is really, really valuable. So that is the first and most important. The second is the products that we build on our data platform, like the one I was saying earlier, Raya, which basically is an AI-based QC system that we've spent like a lot of R&D on making it really accurate in terms of being able to assess data quality and therefore reduce average handling time on the pipelines. And then there's also this other product that's very similar that we call Flow, which is essentially ultimately going to allow us to build like a cursor-like data platform experience where a lot of the rubrics and other things are generated automatically, not entirely, but at least, you know, some portions of it, which then experts kind of review. So it makes their job faster. And as tasks get longer and longer horizon on these data pipelines, we must do things like that to try to control the kind of average handling time. So that's also an important move for us.

Cool.

So I think we had a question. Hey, Gabriel, did you wanna jump on here?

I guess like I had maybe 2 questions. Like one is if there's any, like the one main piece of advice you would give to like a traditional like marketplace, like founder building in this day and age. And then I wonder if there's any huge adjacent opportunity like similar to a space you're in that you feel it's only gonna get bigger and bigger that is potentially being overlooked?

I think in terms of advice for marketplace buildouts, I would say the thing that comes to mind is, and you know, this is a bit cheesy to say, I think a lot of folks have sort of said this, but I think it's important, which is you should really build your marketplace In a way that as LLMs improve, the mechanisms within your marketplace improves. Like even if you stopped development, I think that the test is like, if you pause all development and you just kind of watch LLMs improve, will your marketplace and the quality of your product improve over time? And the answer should be yes. And I think you have to kind of take that, you have to take that to— almost an extreme where if there's certain things that don't actually require LLM usage, you should still consider LLM usage. And then you sort of will pull back on things that actually end up not requiring it. So this is one thing that we have at MicroOne, which is we try to use foundational models that we, you know, that we ironically help improve as well, but we try to use them everywhere in our products. And of course, there's a lot of parts where we do have to build smaller models on our end, like our proctoring model is an example that has nothing to do with LLMs. And we just kind of build a traditional ML model from scratch. And there's a bunch of examples like that, but we sort of bias towards foundational models first. And I think this is a very important kind of product architecture decision that I would recommend. And then, In terms of adjacent areas, I think this area of enterprise evaluations is very much underlooked. I mean, you know, we have a product that we actually just launched in this space. So we're, yeah, this is a bit competitive to us if you go in the space, but I think it's okay 'cause there's a lot of room here and there's not many people actually building in the right way in the space. And the space is essentially helping enterprises kind of rethink their product development lifecycle. And what I mean by that is if you think of like deterministic software, you're, the way you test deterministic software is you just, you have a QA engineer and they sort of look at whether or not things are working and it's a very binary thing. If the thing works, you can go to production. But for probabilistic software, or in other words, agents, you have to take, you have to take the action space of that agent. What are the things that it can do? And then you have to qualitatively assess each of those actions by domain experts that know how to do those actions and see how well the agent does on each of those. And this is something that's not well understood by enterprises yet. There's always this notion of like, okay, I'll run evals at the end. as like a final testing before I go to production, but it's actually the wrong way to think about it. You have to actually, you have to eval your way to production. Like you have to start the product development with evaluations. It's really the, one of the only things you need to do actually to build agents. So building in this space, I think is a very important opportunity that there's many things to do in the space. One of them is obviously what we do, which is like providing the human experts and data pipelines to actually create the evals. But there's many other things that are adjacent to build in this space. I think it's one that will be one of the largest markets that will exist that does not exist yet.

It's really time, you know, Ollie, but huge thanks for taking the time to join us here for the, for the group chat. I think this is, you know, fun. I know we bounce around a little bit, but, you know, covered quite a bit. And it's great to hear about all of your experience, you know, learnings and, you know, appreciate, you know, helping answer some of the, some of the questions that we had and actually had, you know, my usual kind of closing question that I want to ask. And it's somewhat related to Gabrielle's question, which is, you know, if you could go right back to before you started on MicroOne, what would you tell yourself about marketplaces specifically?

don't underestimate the supply side. Once the supply side is there, demand often comes in some form. And this is what we experienced, you know, a year, year and a half ago where we sort of made one fundamental change that resulted in our supply side really exploding. And then that's when the labs actually became customers. All of them became customers. And Of course, there, there's like, you can point to many other things that result— resulted in labs being customers, but, but fundamentally they're there because the experts are there. So like, I think oftentimes this, um, oftentimes the focus on supply side gets a bit ignored and, and, and, you know, there's only kind of focus to generate demand, which of course you need to balance. That's the whole kind of paradox of marketplaces. But, but I think the supply side being emphasized in a lot of cases is, uh, is actually Possibly more important than demand.

Yeah, great parting advice for us. And then last but not least, time for a quick plug. Where can we keep up to date on Uphold along?

Yeah, just my Twitter, I think would be good.

Awesome.

I'll include links and have lots of marketplace founders coming your way. So once again, I really appreciate, you know, taking time to join us for the group chat. And congrats on all the progress.