Your AI Architecture Is Wrong | Agentic EP 05

We're going through a massive platform shift and probably much larger than all of the ones that came before it. Forbes just came out with an article. New York Times just came out with an article. The agentic is the next big thing that can act independently. It can be given goals, objectives, projects and complete tasks with minimal human intervention.

A really common mistake that LLM builders do is that they over rely on LLM tool calls and like what people will observe is constrained decoding, which is really how people perform these tool calls tends to negatively impact the ability of the LLM to perform a new type.

Of design choice that pops up to riding the line between too restrictive and too general and figuring that line out for your specific enterprise task that you're doing is probably one of the most important engineering challenges. Slash piece of information that you have to figure out.

What I like to do from a go to market perspective is like organic partner, paid, get on the phones, hear the pain points, fold that into the marketing message and then when that's working, then we layer on things like paid traffic, content marketing, et cetera. If you have the best quality solution in the world, but there's no trust in it and there's no acceptance, you're not going to get the outcome.

Welcome to the Agentic Podcast where we dive deep with the pioneer AI founders reimagining business and investors betting on them. We bring you the war stories, insights and perspectives you won't hear anywhere else. I'm your host George Van Darrien, the.

General partner of Untapped Ventures, a VC.

Fund that invests in early stage AI startups. Let's go.

Hello and welcome back to the Gentec Podcast. So in episode three we heard from six incredible founders in San Francisco and today we're continuing that conversation from the heart of LA Tech Week at USC, where we brought together three founders, including Daily AI, TenSaltacks and Nullify, sharing their fresh perspectives on building, scaling and innovating in the rapidly evolving agentic AI landscape. This was recorded live at our Gentic and Vertical AI event in LA at my alma mater usc, and this discussion dives deeper into the essential questions around how AI startups are staying competitive and innovative in a constantly evolving landscape. What key trends are fueling the rise in agentic AI, how founders can navigate real world application of AI, and what it takes to sustain growth and fundraise in a market like this. So if you've been following our series, this episode is going to give you even more insights into building a successful AI venture. So with that, let's dive Right in.

We're going through a massive platform shift, probably the largest of our lifetime and probably much larger than all of the ones that came before it. It's gonna impact every aspect of our professional and personal lives in the coming decade. And if we take a look at this simplistic timeline of generative AI, you know, first we needed compute, and Nvidia provided that with their gaming and graphics chips that was leveraged by the LLMs like OpenAI and Anthropic. Once that became a thing, then we needed some new infrastructure, right? Some AI infrastructure. And so you had LangChain and others coming out with infra. And then that led to people using the OpenAI APIs and creating horizontal AI apps. And then you had the vertical AI deep domain experts come out and say, you know what, I'm gonna tackle a very specific, specific problem in a very specific industry, like, for example, Harvey and Legal. And now we're in the era of agentic AI on our way to AGI. So agentic is the next big thing. If you look, and I have, if you look at podcasts, if you look at media mentions, nobody was talking about it last year, early this year, it starts popping up in April, May, June. All of a sudden you could start seeing like, Forbes just came out with an article, New York Times just came out with an article. Podcasts are talking about it. We have our thought leadership pieces on it over there on the top, right? It's a five part series there. Um, so what is agentic AI? Agentic AI is a type of AI that can act independently. Um, it can be given goals, objectives, project and complete tasks with minimal human intervention. Um, so the things that makes it different than other AI is the things that I set. So goal orientation, autonomy, adaptability, training. Um, and so Microsoft got us used to this concept of the copilot, right? Their product is called Microsoft Copilot. And so we use this a lot, the Copilot. But with Agentic, we start seeing the, the AI sliding over into the pilot seat, right? So now if we need to, in this case of the airplane, if it needs to grab a flight to San Francisco, the pilot can take care of it. And there could be a human in the loop. But also the systems are getting better. And you'll hear from some of the founders, there will be times and use cases, and as things get better, that for a number of the projects and workflows, we won't need a human in the loop and AI will just be able to do a lot of it. So this is pretty Exciting because it opens up, it's going to have a profound impact on many different business use cases. These are just a sampling from cybersecurity to just business software development. You know, we can give the AI the objective, like right, you know, create this app for me and then it could go design the architecture, write and debug the code, oversee the quality assurance, maybe source some, some users for testing in healthcare, HR finance. You can kind of read each of the bullets. I won't go through it, but pretty profound impact taking things to the next level. So if we go back to this chart and take out the logos, you know, I think in investing, just like in hockey, you have to invest or skate to where the puck is going, right? Not where it is or where it's been. And so in the last couple years we made investments. We have a couple chips companies in our portfolio, we have a couple foundational layers companies infra, horizontal, vertical. But right now I feel we're really in agentic AI on our way to whatever is next, like quantum. We've also already invested in quantum AI and so I think it's a really exciting time to be in agentic AI. Over the course of those 20 years, I called on every industry imaginable, military, government, healthcare, finance, real estate, and every horizontal imaginable like HR automation, AP automation, AR automation, billing automation, integrating with ERP, CRMs and all that stuff. So now I feel like the team at Untapped, we're having a blast. Because it's easy to meet with B2B AI founders and be able to quickly understand what they're building, what problem they're solving, how to go to market, how to sell it, how to market it. I think we can make good decisions on what's a good investment and then be able to be high value add. So hopefully that was helpful. Let's just start off with some intros. Would love to have you guys introduce yourselves to everybody here, your name or your base, your company and you know, go into a little bit about your, your origin story and what you're building.

Yeah, my name is Ari Attar, I'm the co founder and CEO of TenSaltSacks. We're building autonomous data scientists. So specifically starting on the data engineering side of things, you can imagine these autonomous agents brought things into your think cloud environments, picking up your existing data set as is. So let's say you're using data tools like Spark, Airflow, aws, Bash or whatever. They'll drop in, run your environments as is the same way that your engineer did us so this is fundamentally allowing you to horizontally scale your engineering team, which is unbelievably valuable to customers. I got started in the machine learning space back in 2018 when I first graduated high school. I got really interested in the space because of pasla when they were first showing viability of full self driving gear to computer vision. And just learned a lot of fundamentals on my own and was lucky enough to get a job in the ML space just straight out of isil. So I've been in the space since then.

Hey everybody. My name is Joe Stulte. I am the CEO and founder of a company called Daily AI. I'm a five time founder. I've been fortunate enough to have three previous exits including Grow Flow, which we sold for 62 and a half million bucks, and a company called Lottery.com, which went public for over half billion dollars. And my journey in AI really started when I was selling my last company. It was around 2021. And I just had sort of that founder's intuition that there was something really big happening in machine learning. So as the ink was still drying on the acquisition of that company, I started reaching out and trying to figure out who could I connect with that was doing something interesting at the intersection of marketing. And at the time my lens of AI was basically just machine learning. And so I got introduced to who's now my co founders, Evan Pagan and Dr. Peter Diamandez. And they kind of had this thing in Peter's laboratory that hadn't seen the light of day and I joined as CEO and today we're really focused on building AI automated email newsletters that get extraordinary open rates. We work with some of the top thought leaders in the world. We have over 25 New York Times best selling authors on the platform and then 250 something small brands you've probably never heard of in 30 different industries. And what we're really excited about in the future is building one of the first reinforcing learning agents in marketing to help solve very specific challenges around direct marketing for small businesses. We love small business. They are, we believe that they're the driving force of this country. You know, most people, most entrepreneurs are running small businesses. And as the agentic capabilities start to scale up, there's lots and lots of small brands that don't have teams or capabilities in marketing. And so our vision is that basically every small business can market without being a marketer. And we want to fulfill that through AI and agentic AI.

Hey, so I'm Jackie, I'm the head of AI at Nullify and Nullify. We've raised 4.5 mil from investors like Two Sigma Conviction, Seraguo and OIF in order to build an AI application security engineer. And we're focused on mainly helping enterprises provide security on top of their existing applications, performing SaaS scans, DAS scans, et cetera. And so yeah, I guess we're the exact opposite in terms of customer base, but we're very focused on ensuring agentic AI applications work properly and sort of these enterprise environments.

Nice. Perfect. So welcome. Great to have you guys. Um, so, so let's just kind of go back around in the reverse order and maybe go a little bit deeper. Like you said, AI security engineer, you know what that means. But that might have gone really fast for everybody else. So let's kind of go into, let's unpack the, you know, who, who is the customer that, that has the problem? What exactly is the problem they have? And then how is your solution unique in the, in the market?

Yeah, sure. So a really good way to think about enterprise security is just the fact that imagine you had 100 software engineers, all of them pushing out code, and a lot of the time they're pushing out just code in order to ship features, deliver things for their product managers. And a lot of times the code is going to lead to security breaches because it's not been properly scanned or checked and tested. And there's a lot of security flaws of those code. No brainer there. But what's really common in the market is that CEOs always pushing for features and very rarely is there a correct balance of security towards software engineers. Like the ratio is often very, very low. In fact, a lot of our enterprise customers right now it's basically a team of like 1 to 4 and the need is really clear. And a lot of these mid market enterprises really struggle in fact with hiring security engineers. And it's very obvious we've seen top tech talent basically snubbed from basically being able to dominate the hiring and recruitment market for security. So a lot of time the security engineers are understaffed, typically CISO level suites and they really need help. And so where we come in is that we help them provide essentially like AI application security talent. And obviously we're not promising everything, but there are certain things that are vastly enabled with AI and, and we're focused very much on just those niche use cases. And for us in particular we're focused on how do you fix vulnerabilities, how do you triage vulnerabilities and how do you detect These vulnerabilities in the first place, which are common line items that you would find with these application security engineering job descriptions. For us, what we really, really care about is just augmenting the team of these security teams so that they can actually, instead of worrying about, oh, where do I find the next hire? They just focus on making sure our application works and giving us feedback to improve.

Where should I go, Dr. Balderian?

Well, you kind of already talked about it a little bit, so maybe you can now talk a little bit about how it works. Like what is the process for when a customer signs up to it running and what are some of the results that you guys have seen that are industry leading?

Cool. Yeah. And then maybe I'll just start with the problem. Like how many of you have started a business before? Okay, how many of you have done marketing in your business before? How'd that go? 40% of the content you put into the marketplace actually pushes your ideal prospect away from the sale. Right. Because it's intrusive. Wrong message, wrong time, wrong channel, wrong offer. The first of those three are math problems. So as we think about AI and machine learning and the combinatorial factor of generative plus machine learning, marketing, which has traditionally been a one to many activity, is now going to be a one to one activity. Back in the day, David Ogilvie, who's one of the pioneers in marketing back in the Madison Avenue days, would go and knock doors and he learned how to sell first and then he turned one to one selling into one to many. And that's what became advertising and marketing. And so it's funny how everything in its excess becomes its opposite. Now we're going backwards. Now marketing is going to be sub segmented all the way down to the one to one level. So that's kind of our vision as a starting place. And so the way our tool works is really simple, is we take the position that most people don't know what they're doing with marketing and there's different types of risk you can take with your marketing, right? You can buy traffic and you can go on social platforms, you can publish content and play again the social platform game. But in that world, you really don't have any ownership over your data. Right. If Facebook decides to make a pivot or Google decides to change the algorithm on any of their platforms, all the hard work you put into building your content could implode overnight. So we wanted to focus on the kind of marketing channels that were owned, right? So when you own somebody's SMS list or Their email, right? If the platform starts playing games, you can take your toys and go play with somebody else. And the way that our platform works is really simple. We've built a series of really small agents that help solve relatively complicated problems in marketing. And the first one we started with was email newsletters. Because when we talked to different brands and I polled a bunch of people, I realized a lot, newsletters were getting hot, right? You see things like Milk Road and the Hustle and even Morning Brew. A lot of these companies were like, newsletters were sort of having a renaissance, which I thought was interesting. I also looked at some of the data and I figured out that email has more daily active users than any single social media platform in the world. So it's kind of like the original killer app. So we thought, what a fun place to go and innovate. And so when we looked at the data, we found out that people were spending between two to three weeks, not days or hours, but two, three weeks, to write a single marketing piece. And I thought, well, we have Jasper at the time and all these generative tools, like, what's the problem here? So what we tried to do is basically just collapse that whole thing. And we realized that setting up a newsletter is complicated. There's a lot of micro decisions you have to make. They're setting up DNS records. So you basically set up a series of agents that make that really simple, right? You scan your domain and if your website, your business website is constructed somewhat coherently, we can derive your target prospect, your markets, your products, what they're likely to be interested in. So we just made it really easy. We launched this. And the best part about it is that every time one of the subscribers on your newsletter opens it, we have machine learning running in the background. So it gets smarter and smarter and smarter. We think about them as interest letters, not newsletters, because they're based on what your people are actually interested in. As a marketer, I like to say there's three forms of truth. There's what the brand thinks the market wants, what the market says they want, and that's what they actually click on. We try to over index on that third one so that we're actually giving them more of what they want, thus keeping their attention, which is arguably the most precious commodity in the world, attention. And so that's really the process that we've started with. And newsletters wasn't like we didn't want to be a newsletter company, but it was a great way to monetize. We launched in the middle of a tough kind of tech season. The end of 2021, beginning of 22 and we were able to get pretty quickly to 1.2 million in ARR in about 14 months. But that's not really what's important is that we got over 1 billion interactions recorded on our platform. And so because we want to create a self reinforcing platform that learns the more that we go out and do this in the world, that was really what we prized the most. So that's kind of the problem we're trying to solve and how we're thinking about doing that in the future is going out beyond newsletters and going into other things like SMS and WhatsApp so that brands can send the right message at the right time. That's not intrusive to somebody that actually wants to get it when they want to get it on own media channels which the brand actually controls.

Yeah. So I'll start with the core thesis behind our business, which is that if you look at the broader kind of technology vertical, companies that start very data and data science heavy end up dominating in whatever market that they're in, right? And the big question here becomes, okay, why doesn't every company out there just build a very heavy data science team and build complex algorithms after their applications and you know, build the same level of technology that a Facebook or by dance of builds, right? And the answer is actually very simple. It's human capital, right? For every 250 software engineers out there, there's only one data scientist. And this fundamentally makes it very difficult for most companies out there to build a scalable data science organ house. That's the core problem that we're tab with now. There's a lot of technical challenges there of okay, how can you solve such a technical space like data science through autonomous agents which we've seen, you know, pollution all the time get sucked every loose. And two, how can you continue to wonder existing stacks as is? Because that's, that's a unique value proposition that we're offering coming into our environments and running them the exact same way that this small data science team that you have in hubs is already doing to do this. We've spent a lot of time on like the basic genetic frameworks themselves and we built one for track for this. We'll talk more about this on one of the technical questions. And a second, almost like a middle layer between these agentic LLMs and the base infrastructure that a lot of companies have in place, which we internally call an LLM compiler. It basically takes in a Lot of configurations and inputs from the LLMs. And the same way that a normal compiler compiles your C code into 0 to ones, this layer converts that elements outputs into machine code, which is actually what you know, the base infrastructure APIs typically expect. And we found this works extremely well in terms of driving a very high degree of reliability in these hyper complex environments and data science workflows. So we'll get more to this on more of the technical questions.

But yeah, actually that's, that's where I was gonna go next. So the, the last session we didn't get a chance to do two things. One is like take Q and A. So start thinking about what questions you want to ask. And then also we didn't go as deep technically, so let's do that now. So why don't you just kind of continue where you left off and we'll go back around the other way and let's get into for the ones that are building in agentic and they're starting out, you know, let's talk about process tools people like, how did you build it? What did you learn, what went well, what didn't.

Yeah. So I often find that it's helpful to actually take a step back and try to think about and understand how LLMs actually work the ends, right? Just like any other deep learning model, they're learning some probability distribution of some data that they will pre train and find right now if we think about the performance of LLMs as almost like a decay curve where the Y axis is a performance on some say enterprise specific agentic cost, right? And the X axis is a complexity or an entropy in the costs that you're getting as NA days, right? As a vertical ejected company, the true value that you will be driving the customers will be on those more out of distribution tasks that general agents cannot do. Right now the question becomes, okay, how do we solve for this, right? If fundamentally alums are learning some distribution of data, we can either take one approach which is getting a bunch of synthetic training data and teaching them how to do your enterprise specific paths, which is very complex to do, right? You need to generate a bunch of valid syntactic training data of hey, let's say we're building like a sales office agent. How do I go to Salesforce specifically put in weeds the exact same way that my sales ops team does right now that is very difficult to get synthetic info. Or we can go the complete opposite direction and try and think about building a more abstract representation of our past that is in distribution with what the Elements are already good. And I'll give you guys an example, right? So specific to us, one of the primary ways that. One of the most simple ways that engineers run data science and data engineering jobs on, let's say like cog, like aws, is using Python and AWS bash, right? To do this, there's so many steps that you have to go through. You have to create the iPhone code, you have to containerize it, you have to push the container to the container registry, you have to create a backdrop config and so on, right? So we can either get these agents to force them to use the base aws SDKs and APIs and do each one of those steps and we have to make sure that they do everything accurately. Or you can go the complete opposite direction like you said, and almost go the middle layer, which acts as a translator between the machine code, which APIs AWS base APIs, and the outputs and the burst that we're taking here is these outlets output a configuration on the size of the bashdrop in this example. And that compiler layer just translates it into all of the steps that need to happen misorgunistically. And we found that this works extremely well. There's a lot of other innovation that we've done on like basic depthic frameworks as well. So you're using reward models, which are becoming really popular in terms of ranking the elements outputs on what is accurate, what is not, and redoing certain steps if the reward model gives it a really low score. But the core innovation here is this middle layer that almost translates the LLM outputs into what these base APIs typically expect. Any determination support?

Nice. Can you follow up on that?

Yeah, I mean, I'm not my cto, so I'll give it from my perspective like, like my, my hype, what we like to do. What I like to do as a founder is like, I think the difference between a founder and like a really high agency, high performing, like team member is like founders have the ability to create a reality that doesn't exist and pour so much certainty into that that VCs and team members believe it too. And one of the things I have a fairly high conviction rate on is that software is on its way out, right? Like my model of the future is more like C3PO. Today it's text and chatbots and voice, just more and more. Pretty soon it's going to be video and then we're going to stuff it into humanoid robots, right? And so if you think about that the way I think about software and Screens and the user experience is sort of like back in the day, if you go change your DNS record, you get a little warning that's like, hold on, get a real tech person in here. Do you know what you're doing? You might break it. Like, I sort of feel like the way that we interact with software today, that's how it's going to feel to future generations. And you're going to have more and more agents in the background working on things where a human doesn't need to see the screen or to see the interaction, where it's appropriate to see something visual, like a slide or a graph or a video or something like that. Definitely we're going to have that interaction. But so that's kind of the design principle that I give to our product team. How do we create things for that future? Assuming that it's going to exist? I couldn't tell you if that's going to be three or five or six or seven years, but certainly in the next, I don't know, decade or so, that's where I think things are going. And so we sort of work backwards from that. And I like to think about agentic models as like little humans. Like when I hire an intern, I don't go and say, hey, I'm going to go let you basically do whatever you want and ship code or ship product or go and like another human with more experience of the experience curve is going to go take a look at that. And so what we try to do is compartmentalize their work down to the smallest unit of work. And then we let them try because they're high agency people. And then if they don't figure it out, then we're gonna coach them until they get it right. That's the same exact model we use in our company. Right. We just break each workflow down into its, like little specific components. And then we build micro agents that handle those. And once they've been tested on enough data and we have enough confidence and we've, you know, we remove the human from the middle and all of a sudden that little agent now performs admirably well. So when we first started, like I would say about 8 out of the 10 newsletters that we created needed some kind of human interaction. Now out of 100, maybe two need a human to touch it or look at it or make some kind of edit. And so we just break these things down into really small chunks. But what's cool is as these LLMs get more capable and even how they train the models are changing Right. Like if you're looking at the new Zero1 model that's got deep reasoning, it's using inference. So they're basically applying the same logic. They're going out and saying, hey, I'm going to watch how people interact with the tool and we're going to figure out when it works. Then we'll double down on that as like a mini use case that we can use to further our existing training. And so that's the same way that we're approaching building our product. We just try to break them down into sub little baby agents, train them on a very narrow scope and then let them graduate. And then again, our thesis is that if we can nail the core use case that I've described around direct marketing, that we can use the same model to scale and go further and further. But as a multiple time founder, I've learned the hard way. Innovation is best described as make it work small and then go big. Every time I've tried to make it go big, I end up, you know, it's kind of like when you touch the stove as a kid. It's like, how many times are you going to do that before you learn? Your hand gets burned. So my experience as a founder has just been like, how can we make it work small and then scale it up and how do we do that as fast as possible with as much interaction and data from our end users as possible and also how they feel about it. The last thing I'll say is this. A lot of the stuff that we've been wanting to do that we're now finally doing this year. Like we wanted to send, for example, personalized newsletters where every subscriber gets their own edition. But when we tried to sell that to the marketplace, guess how many people wanted to buy that? 0. Right? So I have this equation in my head. It's the quality of your solution times the acceptance equals the outcome. Right? And so if you have the best quality solution in the world, but there's no trust in it and there's no acceptance, you're not going to get the outcome. It's like if my grandma was still alive and I handed her an iPhone and said, call 911, no call would get made. But if she pulled up the old rotary phone, then we'd get a call made. And so you can. Even as we interact with more and more brands and customers and people in the marketplace, we're just noticing that the trust is going up. They trust us with more things, they trust AI with more use cases. So it's sort of this interesting dance between where's the capability at in terms of large language learning models and AI overall and where's the acceptance of the marketplace. And what I've observed is there's a real convergence between those two things happening. So it's just as much as we can unlock technical capability and break that down to its gnarly details, really it comes down to like the. Is the everyday average person or company going to be buying this? And again, like, we can get into this later, but the way that we earn their trust is we have a mantra in our company, like our number one company value is outcomes over outputs, right? We want outcomes, we want the click, we want the meeting booked, we want the sale. We don't really care about the output. The output is a great leading indicator. But where everybody else is sort of getting high on the novelty of cute outputs, that doesn't really make any difference if you're a small business owner and it doesn't actually move the needle in your business. So the more that we've been able to earn trust on moving the needle on specific outcomes for the business instead of focusing on outputs, the farther and faster we've been able to go and unpack those things. So hopefully that's helpful.

That's great. And we'll talk a little bit more about like general lessons learned and advice to founders, but that's a good one there. All right, Jackie, what are your thoughts on this?

We made a lot of lot of mistakes when we were starting out and I think right now everyone is actually figuring it out as it comes. What we've seen in the market is that there's probably only about eight startups that are actually productionized agents. So just to restate that number, it's like way less than 10. It's like eight. And what that means is we're pushing really the extent of what's currently possible and for a lot of what people are currently building. There's a lot of misunderstandings about LLMs and how people are building should be building on top of these tools. For example, a really common mistake that LLM builders do is that they over rely on LLM tool calls. And what we have observed and what a lot of people will observe is constrained decoding, which is really how people perform these tool calls tends to negatively impact the ability of the LLM to perform. And so when I translate that to what that means is that whenever you use tool calling, which is an API, you're really reducing the performance of your LLMs and its base knowledge. And we've been very deliberate with how we program these LLMs and agents. And we've also decided to make investments in this space that are quite counterintuitive as a result. So for example, a lot of companies right now are like, oh, evals, evals, evals. And what I've seen would love to hear other people's opinion as well, is you measure in, your evals becomes what you optimize. And with Swebench, for example, you're really optimizing open source issues on GitHub. And what you quickly realize is that if you're building in the coding agent space, enterprise software does not match open source Python repositories and you're basically not really making the right changes. Usually they're not even the right language. And so if you're optimized on swedbench, awesome. Go to market, terrible predictor of performance of your product. And there's so much noise and sort of what we've seen is the LLM investment space is that there tends to be three types of companies building in this space. You have AI, agent labs like Devin, like Sakana, AI that are building things on top of LLMs. You have model based performers. So you've seen magic.dev, poolside AI. These are training foundation models. And then you have a third type, which is what we call internally as Agent OS, which is application of LLMs within customer environments to ensure optimal performance within that particular key. And sort of what's been so correctly mentioned before is how can you actually model the probability distribution so that these LLMs can work in a space that's set up to succeed? And that's like a boundary that's very fine. You have to be very careful about where you devote resources. And so if you're a founder who's raised in this space, definitely think twice about how you would build that data set, how you would then productionize those agents. Because yeah, it's like you sort of have to walk that fine line. And we've made mistakes over investing in certain areas. Yeah, we've been very careful to just only focus on certain tasks and sort of where these LLMs excel extraordinarily right now is an information synthesis. So focusing our efforts in those particular areas.

One other thing I wanted to add to this as well, for people that are building in the vertical agentic space that I just want to add to is this whole like middle layer, whatever you want to end up calling it, there's a new type of design choice that pops up, which is almost riding the line between too restrictive and too general. And figuring that line out for your specific enterprise task that you're doing is probably one of the most important engineering challenges. Slash piece of information that you have to figure out. So the same way that you know, like a programming language doesn't have too many abstractions but is not, you know, too restrictive, that's the type of way to be thinking about this and getting that right kind of balance is honestly one of the most important things to get this right.

Okay, awesome. That's a great pro tip there. So I want to move to two exciting topics as well. So let's do customer acquisition, right? Go to market and then fundraising and then we'll go to Q and A. So with that maybe we can start with Joe. And again for both the, the traction question and the fundraising question, share as much as you're willing to share. Can you give us a sense of like, you know, where you're at in terms of customers ARR and like what were the lessons learned? How did you get the first 10 customers and did that evolve over time and what advice would you have on customer acquisition to founders that are just getting started?

Cool.

I mean just right off the bat like we're sales led company. I love sales and marketing if it wasn't obvious so. But I also love building great products. I don't like selling things that aren't great. From a Traction perspective, what we did was super simple. We never raised like a ton of money. We raised it like in smaller chunks at a time just because we wanted to validate it. And as a multiple time founder, there's like nothing worse than like raising a boatload of money and then being like sorry, we failed but. And it's also nothing better than winning, right? So what we did is we got super creative with traction. We found really big partners. So fortunately one of my business partners is a multiple New York Times best selling author and thought leader. So we launched our product to his market first and we de risk it a ton there. We did our alpha product and do we learned a lot. Like the thing that we had back in 2022 is like I'm deeply embarrassed by it that but I'm proud of the fact that I'm embarrassed by that because we went out to market and I was on the phones personally selling it to people one to one before we had a real product to give them to be quite frank. And, and that was how I did customer development in the early days. But that model worked so well. That I went out to other thought leaders that had like, they had my target market and I would build their newsletter. We would like, for example, we built a newsletter for Chris Voss, who's one of the world's top experts in negotiation. And we got him like a 90% open rate, like every edition that we sent out. And then he started sort of, you could say referring or promoting. I like to call it prescribing the way a doctor prescribes medication. He would prescribe us to his people and then we would go find another thought leader and another thought leader. So there's a little mini network effect. The more thought leaders you brought on, we got them a great result. Then they would promote us to their audience. And that worked really well. And we did all that without any paid ads. Like, we didn't really go to market with paid traffic. So what I like to do from a go to market perspective is like organic, partner, paid. So like hand to hand combat, get on the phones, hear the pain points, fold that into the marketing message. Then I'll go out to partners and do what I just described. And then when that's working, then we layer on things like paid traffic, content marketing, et cetera. So that way if you hear about us and you're outside of my ecosystem at least we're doing things like retargeting. So when they land on my landing page, we're getting a lot of traffic to that and then we're retargeting to get them to come back to the page. And we split tested our landing page like a ton. And that's how I came up with our like, each word on our landing page is incredibly optimized. And that worked really well. You know, Today we're like five customers shy of 2 million in arrival. We expect to hit about three by the end of the year with huge promo coming in the next couple months. And from a go to market perspective, that's, that's the way that we approached it. And once the next iteration of our product comes out, we'll be a lot more aggressive on ads. We're just about to close our Series A, so we'll have some dry powder to go pursue that in a way that we have it in the past and very similar to that model. That's exactly how we thought about fundraising. Right? So this, I don't think this was the kind of project that like Tier 1 VCs were super interested in. I don't know why. But why, I suspect why. But my observation is like, I've raised money for other projects that I think are way more boring than this. Way more easily, partially was the nature of fundraising. You know, from like 2010 to 2019 was like helicopter money everywhere. But also part of it was, you know, just right fit project. Right. You have the right leading indicators, you get checks. But there's something about like SMB marketing and go to market tools that is a massive allergy. I don't think a lot of funds understand it and it's not meant to like demean the capital. There's important partners they don't understand the same way we do. So long story longer. We didn't really go after tier ones. I mean, we had some really great conversations and they were like, we totally see what you're doing. And I could never bring this to my investment board. So we did the same thing. I made a list of all the funds I knew, all the family offices I knew. We ranked them based on capital, the ability to help us out on go to market, the ability to introduce us to great talent. And we were very picky about who we let on. You know, we were very fortunate to meet George early on and that was helpful. We also met Dave Blunden, who's had an AI fund for a very long time and is a deep partner at the AI MIT lab. That helped us get access to some data that really advanced our company. So I did the same thing. We made a list, we made it work small and then we went big. But we're really focused on VCs and capital partners that could really add something beyond the money. So go to market data, deep advice, looking under the hood and really helping us fold time so we don't make silly mistakes. So hopefully that's helpful.

Yeah. Full disclosure. So daily AI was one of my angel investments before I launched the fund. And then we invested once through the fund and then once again. So it's been a great ride with Joe and it's always great hearing the story. Awesome. So why don't we go to Aria next?

Yeah. So from a Germarket standpoint, we've honestly been trying to figure out what is like the optimal approach to selling vertical agents. Obviously over selling you the highly technical and it's a developer tool. So so far things that we found works is the traditional kind of developer tool, bottoms up and then running a top down on top of the bottoms up motion works pretty well like product growth, get it in the hands of the developer and then go from cotton to executives and kind of run an extension play across multiple other teams that so Far seems to work pretty well. We've gotten one really big Fortune 500 company as a customer, which is awesome. A massive gub agency and like a ton of other smaller mid market companies. You guys just launched this here, right? Yeah. So the expansion has been Quaker and I think there's a lot of tailwinds that are playing into our favor as well. Yeah, like honestly the economy as a part of it. Right. People are trying to run leaner, get more outcomes than they were previously with the same lean team, which works really well into our favor and the value that we're delivering to customers. There's also the tailwinds of C suites pushing AI tools, Vanderbilt routes use AI everywhere and obviously this is one of again the most mission critical space that you can use it for. And so far it's been super valuable to them. Fundraising.

Yeah.

Yeah. So yeah, we did a pre seed round back in December with a fund called V Partners out of San Francisco back in December. So now we're doing our seed round. I've talked to a lot of different investors. Some get it, some don't. And this is one of the challenges with building a hyper technical product is a lot of investors won't be able to truly differentiate, you know, two different types of products that seem like they're in a similar space, even though fundamentally they're not. And that is totally okay. So I think it's a big question of like finding the right investors, partners and fund that are highly close to the problem space that you're solving for and working closely with them. So George here again.

Awesome. Thanks Arya. All right, Jackie.

I can't go too much into our go to market fundraising. We've actually heard very, yeah, like very pointed questions from investors actually. The two main ones being does it work and what are people saying about this? So really it doesn't work. Like that's been pretty much it because almost like obviously the agent dictam is like really, really high. Any vertical SaaS to vertical agent, huge deal size, especially because you're replacing knowledge workers which very, very expensive on payroll. The unfortunate bit is almost no one has it working, so investors are very scarred by it and our series A is contingent on that. So yeah.

Okay, awesome. So my, my team's like asked them a controversial question. We, we, we had the vertical versus horizontal, but I didn't think it was as interesting on the founder panel. I'll definitely be asking the VC panel in a minute, but we have time for just a few questions. So let me grab a mic. Does anybody have A juicy question that you think would be like, applicable for everybody versus like a one on one.

Thank you all for that was really super interesting. I wanted to follow up on something you said, Joe, about your equation of quality times acceptance. Assuming you have quality, how do you drive acceptance? I don't think there's any one right way I can just tell you my lived experience. I think part of it is trust. And the way that we've approached it is like getting base hits instead of trying to hit grand slams. So faithful with little, faithful with much. So if I can show up and get you a micro outcome, that'll build a little bit of trust. A little bit of trust that doesn't get you as far as, let's say, like, you don't know. ChatGPT launching and becoming like the fastest, most adopted product in the history of our species. Like, that helped a ton for acceptance on our side. So there's sort of the tide that goes in and out and then I think there's the things that you have that are more within your control. And I think, again, like, there's so much. There's a gap between the hype and where the rubber meets the road. With AI, it leaves a giant space for someone to come in and tell the truth. So anytime there's a massive distortion in the marketplace, it leaves a great opportunity for you to come in and tell the truth. Even if that means, like, you can't promise a bigger promise. I think there's a certain subset of the marketplace, like a higher LTV stick around with you for a long time, deal with your alphas and your betas that will come and play ball with you if you do that appropriately? And that's been my experience is like, hey, how can we call it the distortions? How can we be like, brutally honest about where we fall short, but also be brutally and ruthlessly committed to where we can deliver value and just, you know, make it work small and then go big. Like I said, like, that's been my approach. I think the other way to do it is like, you know, you just look at like blockchain, right? Just keep or elon, you know, like just keep saying really big promises and really paying the vision. And the vision gets bigger and bigger and bigger and slowly the capability catches up. But the more I say it and wave my hands fast, the more you start to believe it. I mean, there's something to that, right? Like, perception is reality. I'm not, I'm not trying to knock the greatest capitalist entrepreneur that's ever lived and recorded history. But I'm just saying that's another way to do it. That's just like I think it comes down to preference and you know, your capital. But my preference has always been like hey, let's like let's hit base hits. Let's do great work and earn the trust.

Great panel guys. You mentioned distortion in the market. And vertical AIs are very much competing with SAS businesses all over the market. And I think one of the fundamental distortions is that AI is a deflationary technology driving down price, driving up quality. But all the SaaS competitors are going to be charging more as they bring AI capabilities onto their platform. So how do you guys reconcile this idea of AI as a deflationary technology, but then reconcile that with how you price this in the market and price.

This against your competitors on stock companies as competitors? I think as a vertical ejected company you have some advantages that the incumbents are fundamentally lacking. And this is one of the common questions that cop up on investor conversations is like okay, and like I'll just use us as an example like the data engineering space, right? Why can't an incumbent data agent platform come in and they have to slap some agentic functionality on top of their obligation and just sell it the same way that we do. And the reason for that is fundamentally as an advantage such as being for example fully agnostic and being go from the ground up to your staff. Right? So as a verbal ejected company, the promise that we're getting customers customers is no longer a faster time to an outcome the same way that a SaaS company does, but it's the final outcome itself. That's what there's somebody to buster. And I think that on its own is fundamentally a completely different type of business model that Cook was actually actually be deflationary. But there's the other side of it which is the number of outcomes that you can now get is significantly higher. Higher meaning you in a way are still kind of off balancing that. In a way you could say the same thing about pob, right? POW is in a way deflationary to data centers and price of building software is not significantly lower with Bob, but the amount of total software that is not being generated is s becoming higher. So I think in a way off balance is out because you can now get different and more moderators on the other side.

I mean I think GDP is going to grow and I'm not sure how deflationary it'll be.

I'm curious how you remove humans from your development cycle loop in terms of evaluation. It feels extremely necessary, especially in the beginning, but also probably the largest bottleneck for developing more quickly. So I'm curious to hear how each of you and three very different products work towards removing human evaluators from your development flow.

Yeah, so right now what's been really critical for us is just like making sure it's not like a basic human evaluator, but it's built into the UX and it's like a semi automated way of capturing observability data and then linking it back to how people basically would use your product. So for example, with customer support agents using that as an example, you would run into an edge case, it sucks, you thumbs it down, people should know. And so that's been a good way to start, but it almost has to vary by task. And I'm hesitant to give a conclusive answer on every single task. I think that you have to be very careful with the agentic UX and how you measure confidence. And that's a whole separate discussion. We have a whole research, we have someone dedicated to research in that.

So yeah, I mean ours is pretty easy. We just hired a bunch of humans to fix everything and every time they fixed something we recorded it and we were very pragmatic about it. And then I had this hypothesis that there's a threshold with which we make the customer the human in the middle. And I think we've reached that. Right. So we've now like about a month ago started letting people do self service again because our hit rate on article curation and newsletters is high enough. And now we're just very transparent. It's like, hey, it's not completely autonomous, you're going to have to make some updates. But when they make updates, we allow them to make comments and we're making inferences around that so that not just around what they change, but what the people actually click on. So that was our model because first we did self serve and it flopped, didn't do very well. And then we went to completely done for you and we just did it ourselves. And then I was like, man, if we can get this to a certain threshold, then we can make the customer the human in the middle. And then after enough data points, when we start getting more and more people on not being, not regulating our scale by our own human hands or whatever, human minds, we're gonna get a lot more data points and fill that gap a lot faster. And that's like kind of final mile for us. But again, we have to repeat that for each use case Right. And so but that was our approach to it, right? Crawl, walk, run.

So I, I think it's two things. One, on like the base agentic frameworks that we built, we added this like critique slash evaluator model that literally just low set the task that the main user is doing and it has the ability to give it update and instructions its adapters. And we found this works really well in terms of like just catching dumb mistakes that the main agent was making and making. Go back, redo that step if it doesn't do something right. So that's one part. Second is it's highly dependent like Jackson said, on like the task itself, right? Maybe like a right action on your whole data warehouse. Do you even actually want full autonomy?

Right?

Like does that, you know, small level of extra autonomy on that specific step actually add that much value? And in most cases the answer is actually no. And customer would prefer to have the peace of mind of having a human loose step on important. Right actions just so they, you know, don't break everything in, in their, in their existing infrastructure. So it's highly fast to Bennett.

Let's give them a round of applause. Good job. Great job, bro.