Latest / The AI Ecosystem Podcast / Introducing Trace & Building AI Native Products from Scratch

Transcript

Introducing Trace & Building AI Native Products from Scratch · The AI Ecosystem Podcast

• Bob Moore: Hey everybody, welcome back to the AI Ecosystem podcast. We love doing this thing. It's so much fun, and we appreciate everybody who listens, sends in feedback, and shares these episodes. we are starting to talk about them more and promote them more, which is really fun now that we've got kind of a core 10. We wanted to make sure that we would actually do this for more than 10 episodes. It looks like we're sticking with it, which is really great. So we've got a bunch of new listeners and followers, and for folks that Lindsey DeFalco: I think so, yeah. Bob Moore: Care to give us five stars or a big thumbs up on whatever app they listen to, we finally are ready to ask for that. So we do appreciate it. today, you know, we realize that many of our most popular episodes are the ones where we talk about real stuff, real applications of AI, the way real companies are bringing it to life. And we have a doozy for you today, which is that it over the last seven months. I and a very small team of folks at Crossbeam, in addition to our day jobs, have had a little bit of a side hustle, which is building and launching an entirely new product line for Crossbeam, which is gonna go live on August 1st publicly here. But if you're a listener to the podcast, we're gonna let you in on this little secret just a little bit early as a thank you. and we wanna spend today digging into the full life cycle, this six-month sprint. That has led to us bringing an entirely new product live and to market. So we'll talk about what it is, where the idea came from, and how it is that we used AI to make this thing absolutely bonkers good at a speed and velocity that's unlike anything that I've ever seen in my 20 years of building companies in in tech. so I'm pumped for it. are you pumped for it, Lindsay? Lindsey DeFalco: yeah, yeah. And I think this the specific thing we're gonna talk about is is interesting in how it came about because it is both born out of being really, really complementary in terms of Crossbeam and what Crossbeam does, but then was also born out of a desire at Crossbeam for us to actually build an AI native product and see what that looked like from the ground up. you know, not working in like a legacy non-AI native code base. Bob Moore: Mm-hmm. Lindsey DeFalco: and so it's this like beautiful union of a really, really nice use case that's also gonna benefit everyone that uses Crossbeam and this this internal goal of seeing what like best case can look like for for building AI products. So yeah, super, super excited to to get into it. And I know you have been spending how how much of your time on this? This is a is a big big effort from you. over the past three months. Bob Moore: Yeah, it's you know what's funny? And this is a total it's almost like a meme of the random tech bro who discovers clawed code and suddenly, you know, becomes obsessed with it or gets AI psychosis. I am definitely spending tens of hours per week on this. What's weird about it is that it's all happening like at night. And it's one of these things where by hours worked, it's probably half. It's probably about half the time that I spend at Crossroom right now. I I spend working on on trace, which is the name of this this new product. And however, I think it has expanded the number of hours that I work by quite a bit because I just can't stop thinking about it. and I do think there's a little bit of this this thing where A, it's really fascinating. It's I've stopped playing video games. Like the the dopamine rush from playing video games, which I, you know, occasionally get hooked into, has been totally replaced by this. And There's also this really weird psychological effect of if I don't use up all of my token allocation for the six-hour period that it is right now, I feel like I'm losing money. Like if I don't get physically stopped by Claude because I've used up all my tokens, then I feel like I've failed. and that has resulted actually, ironically, a lot of the most innovative and interesting things about this product, which are these like sub-agent loops and these automated kind of agentic processes that Run that are constantly working to cultivate the the data that sits behind this product. They're all byproducts of me trying to use all my marginally free tokens overnight while I sleep. and it has turned out to be like a really like actually a pretty effective way to build certain parts of this product. So yeah, it is it's a it's a big, big chunk of of this is the equivalent of like founder, you know, takes some LSD and goes off in the woods and comes back with some. Crazy new product idea. minus the LSD and the woods. this is kind of the mode that that I've been in for the last couple quarters here. Lindsey DeFalco: Yeah, nice. so before we get into like what this product actually is, you said Trace, the name of it is Trace, and then get into, you know, how you went about even starting this whole thing and like what it looks like to build in a completely different way. I do think it would be fun to talk about because you talked about cost optimization and we were talking about that last time. And you shared an anecdote with me about like the most recent crazy cost optimization thing that you saw. Do you want to tell people about that? Bob Moore: Yeah. Yeah, I love this. And and I also this is one of the things that I love and I'm mad that I didn't think of it. So with the the the premise of what we talked about last week to save money on tokens or save token consumption, part of that is Tokens are kind of measured on inputs and outputs, but not necessarily on the processing that goes on behind the scenes with those tokens between the time when they go in and when the results come out. except there's like intermediate exchanges between sub agents and stuff that that I'm sure they they dock you for. somebody figured out that The if you have a really long prompt or a bunch of context that you want to provide into a session inside of Claude, particularly with the new Fable models, which is back, Fable 5 is back, baby! with the new Fable models, which consume a ton of a ton of credits, that you can upload an image that contains very, very, very microscopic text. That is able to be parsed and deconstructed and turned into text and then processed as context. And what you're getting charged for on the input token side is less per pixel, per pixel character equivalent, by a massive factor, by like a 5x factor, than actually sending the text itself. So if you have a big chunk of context you always like to send into every session, you can load it into an image that you know Claude knows how to upload and process and pay. One-fifth what you would pay to actually send that context as text. Now you know Anthropic is not actually experiencing better performance on its own back end from this. Like the amount of chip time and storage space and information transfer sh should be greater because you have the OCR and the image deconstruction and all this other stuff like sitting on top. But because of the way that they monetize and kind of instrument billing on top of the various modalities that you can send information in. Lindsey DeFalco: Probably more. Yeah. Bob Moore: In this particular moment in time, and I bet by the time this podcast is out, they'll have patched it. it turns out that yeah, you can basically turn all your all your inputs into images and you can you can save a bunch. And I wonder I didn't see I didn't see anything about this, but it's worth testing if it claw doesn't generate images. Claw doesn't g I was gonna say, could you have it give you your output as text in an image and then have a much cheaper model? Lindsey DeFalco: Yeah. As an image. Yeah. Bob Moore: turn that image back into text for you as like a as like a layer in between and you're like you're getting back to like caveman in, caveman out, only it's image in, image out. But Androbic doesn't doesn't generate images. but I wonder if that hack might work on on Chat GPT or on on Gemini. somebody out there's probably working on it right now. I love it. Yeah. Lindsey DeFalco: Yeah. That's so funny. I'm sure they are. It reminds me of seeing the there's like a tweet about it or whatever and seeing the image with all the text crammed on it reminds me of when you were allowed to bring one page of notes into a test in school. And you would write in like four point font to get, you know, every single piece of information you possibly could from the semester onto this page. it's like the analog Bob Moore: No. It's so great. Lindsey DeFalco: Old school equivalent. all right, cool. So trace, let's do it. So Bob, what is trace? Bob Moore: Yeah, so so let me just start with what it is, which is Trace is a fully AI native, AI derived corpus of data about companies. Really simple. It is not a new product, it is not a new I mean it's new to us, but it's not a new category of product. This stuff has been around forever, for decades. But it really specifically focuses in on being this really reliable, referenceable data set that for any given company out there can tell you what technologies they use. what products they use, what companies they are a customer of, which are all three kind of related categories with little nuances, because a company can be a customer and use three out of the five technologies they use, but having that graph like really solid. Also mapping their partner ecosystem, who they're partnered with, who is partnered with them, and also mapping kind of basic firmographic information, their headquarters, how many employees they have, you know, are they a subsidiary of something else, says it what venture firms have invested in them, that kind of stuff. and it's a really big ass database that contains all that information, but most importantly stays very, very presently up to date with that information and knows how that information is changing over time, which is basically signals around what's going on within a within a particular company. and the interface through which you can access this stuff, not only do you have a traditional UI and the ability to export CSVs and and all those goodies, but it was kind of built AI first and MCP first. So the the mechanism by which you can access this data or agents can access this data is very much kind of built and designed and organized around the premise that it'll primarily be consumed by other AI systems that might then need to use it and route it downstream. That at its core is trace, and it's got a bunch of really cool additional features for monitoring changes and generating lists and you know, managing teams and getting notifications and all kinds of other stuff. But our goal with it started out with a need that we had inside of Crossbeam, that we had looked at many, many times over the years, which is we just wanted to have our own data that we present in the product be enriched by a big corpus of other information that we had. So historically the only thing in anybody's cross meme account is what they know about a company and what their partners have shared with them about a company, which is great. And that's all by the way, all that cross meme data is like over here in one area. And then there's a giant air gap and all the trace data is over here in another area. Trace doesn't touch cross meme, it doesn't talk to cross meme. It has gotten No information from Crossbeam or anybody's Crossbeam stuff. It is completely derived from publicly available data sources and a bunch of really creative ways that we you know, identify and collect those publicly available data sources. But Trace is like a whole other beast. But wouldn't it be cool if, and we've wondered this since before Trace, in Crossbeam, if you don't know how many employees the company has, it's not in your CRM and none of your partners have shared that with you, if we could just give it to you. It's just a p it's a fact, right? It's it's probably Written in the about us section of the job description of that company and in the, you know, whatever, 10,000 other places that it's observable. what if we could just insert that in there? And it's always been a very annoying and expensive proposition for us to do that because we we provide data on so many companies to so many companies that if we tried to go to one of the big data providers out there in the market and we asked for like a blanket right to OEM their data and just put it in our product across millions or tens of millions of companies every single day. They wanted to charge us like hundreds of thousands or millions of dollars for it. And it just never like made real economic sense for us to do it. We figured the only real way to do it would be for us if we could just own that data asset and kind of natively derive it from scratch without assistance from anything, except like kind of starting starting from scratch. So in addition to The firmographic info and things like employee account, the other thing that we always thought would be super cool in Crossbeam would be if there are companies that haven't signed up for Crossbeam yet, and we can figure out who their customers are through publicly available sources, then we could actually have like a Phantom Crossbeam account that's kind of made up of what we could call open data and just have it as though that company's on Crossbeam, even when they're not. So and have it be powered by by these like open data populations or open data partners. And the only thing we would need to do that, again, is the corpus of data that Trace has, right? It's like you cultivate a pretty high confidence list of who is a customer of whom, and then you use that as kind of a this open data partner. And as soon as the real company signs up for cross premium and connects the real data and like gets a real fortified, legitimate, complete data set coming out of their true CRM. You know, we could we can swap that out or or get rid of it or kind of navigate through that. But in the meantime, it gives our network data like this really, real strong and powerful completeness to it, which kind of fills in any holes or any gaps. Where if 80% of the world uses Crossbeam, now all of a sudden 100% of the world's information is on there, at least the publicly available and observable parts of it for that that last 20%. So we wanted to launch this feature. This open data partners feature. And the only way to do it was to to have this corpus of data that Trace is. and then there's a third thing we said, this is like a big if you give a mouse a cookie thing. So it's like, okay, now all the data's enriched with all this cool photographic extras. Now we have these open data partners where you can sign up and you don't even have to partner with other companies. You can just like see the customer list of those companies and account map against them. Now the last leg of the stool is. What if you haven't even connected your own CRM yet? Or there's a thing that's gonna slow you down or you need to go through apps, but you want results today, you wanna see the value of CrossFeam within five minutes. Wouldn't it be cool if we could generate a target list of companies you should care about for you? Basically a look-alike population that says, hey, here's your target market. And what's a really good way to do that? Well, if you have the corpus of data that Trace has, you know who's customers of whom and you know who's partnered with whom. And you can basically draw an effective map on based on how many degrees of separation these customer relationships are and the partnerships between companies and how common it is that companies have a very similar stack of technologies or services side by side. You can create a pretty easy inference model to say, hey, a company just signed up, they have not connected their CRM yet, but we are in a spot where we can tell them or we can identify who very likely a thousand companies that they ought to be targeting are. And we can drop that right into their population. So now you go from a world where a year ago in CrossSpeam, you sign up for CrossBeam and you have to get over these two really big hurdles. The first one is you have to connect your own CRM data. So you can put in your opportunities and put in your prospects, and you can tell us who you care about focusing on. And then you also have to connect with partners, and those partners have to be connected and they have to be willing to partner with you and tell you who their customers are so you can account map. And then finally you get the account mapping. Here's how my prospects overlap with their customers. Well, today, or imminently, you can sign up for Crossbeam, and that entire experience of crossing those two mountains doesn't exist anymore because you can generate a lookalike population immediately where you have a target market list right in front of you that is like here's the companies that we ought to be going after based on a pretty awesome inference model. And you can partner up with a bunch of companies without needing to wait for them to connect their data or connect with you, because what you're comparing the data to is these open data. Partners or open data populations, also based on publicly available data. And then you can look at that account, that beautiful account mapping matrix of like, here's how my prospects relate to your customers across a bevy of potential partners and collaborators, and have a bunch of data that you can act on and start doing the really awesome downstream cross-BME stuff with basically instantaneously. And then when you do connect your real CRM, the data gets way even richer and more robust, and you can track it through your. True opportunity pipeline and your existing customer base. And when you do connect with partners, you get way richer data from them and who owns the account and a truly accurate customer list. And looking deeper into their pipeline, there's still an immense amount of data unlocked that comes from the real data on both sides. But there used to not be a single player mode for Crossbeam. And if we owned the trace data set, the thesis was that we could basically take Crossbeam to like a zero-day, zero-hour. proof of value experience on new people onboarding onto the product, which which felt really, really compelling. so we decided to to try and go out and build that data set. And it felt like the AI era was the perfect time to do it. And I can speak about why, but I'll stop there. Does that make sense or spark any questions about the why? Lindsey DeFalco: Yeah, I think you covered all of it. And what I love about everything you just said is that it touches on at least one topic from I think every podcast that we've done so far, which is which is just a great little full circle moment. So to give like a real example of the open data that you were talking about, you know, we talked about marketplaces. previously, these hyperscalers who, regardless of if they're on Crossbeam or not, maybe they're not sharing their data with you in the way that you want. And I think that's been one of the most popular separate data assets that we see people run alongside Crossbeam is these really large organizations and and trying to get some approximation of their customer list. and so what is so powerful about You know, what you're talking about here and having that alongside all of your other cross beam data is that you know, you can now actually qualify and identify leads that you should be submitting into these marketplaces and they're gonna by nature be higher quality because we know that there's an existing relationship because of this third party data. And so it's it benefits everyone, right? Like that benefits the the hyperscalers in the marketplaces because they're getting higher quality leads. and then it benefits our our customers because they're not wasting time submitting something and then getting a rejection back because there's no no existing relationship. So it's like it's this really beautiful blend of the second party and third party data, which is another thing we've talked a lot about. and I think that's that's my favorite thing about it is, you know, there's there's all the discourse around you have your first party, your second party, your third party data. and it's not really which one is best. It's how to use all of them in concert with each other and where where they're valuable for different motions. And so what Trace is doing is it's really it's like bringing the the second and third party data together in a way where it makes both of them more valuable than either of them are when they stand on their own. and I think yeah you walked through a bunch of the reasons why especially for for people to use crossbeam it's really, really valuable, let alone just, you know, having each of them stand alone and and their own individual value. So Bob Moore: Yeah, it makes makes sense. Something that struck me too during that rant, is most of what I described, in fact all of what I described, was true and could have technically been done before the ChatGPT moment and before LLMs. So the idea that we could collect a giant corpus of data about what companies use what technologies Lindsey DeFalco: Uh-huh. Bob Moore: Yeah, that's been around for a while. the idea that you could build a lookalike population based on the underlying data from there. Yeah, people have been talking about look alike since Facebook ads in 2011. what I think struck us as so novel and the big answer to the why now. Was that historically the companies that have been doing this and have been collecting this data have been doing it in a very painstaking manner that has been very capital intensive and very error prone, and very prone to stale data as well. And what I mean by that is manually writing enormous, highly kind of long tail oriented code bases that scrape data from a bunch of places. Paying tons of data to other kind of third-party contractors who do the scraping for you or do the collection for you. doing various shady things that may violate people's terms and other things that we wanted to stay a hundred miles away from and really haven't done. some companies, if you go back far enough in time that sell firmographic data or technographic enrichment data. you know, have have call centers that literally like call up businesses and are like, hey, are you still located on the Black Horse Bike in South Jersey? You know, like verifying address information and employee accounts by getting people on the phone. and you know these these strategies go go really far back and kind of take a whole bunch of different shapes. But you can imagine how either through engineering cost or through human labor cost, not only does it get extremely expensive and take a really long amount of time to accumulate this kind of data, but it also Becomes very, very difficult to maintain because the the universe of companies that you cover and technologies or pieces of information you might gather just becomes this giant kind of not quite n squared, but like n times order of potential risk, where as soon as something goes stale, what's the frequency with which you are you know, dynamically able to update the knowledge as to what CRM system this company uses or whatever else, right? So, and I think there are companies that have gotten fairly sophisticated with that, but they've had to do it painstakingly. And by writing a lot of manual code and hiring a lot of people and raising a lot of venture capital and I think the the the aha moment around this for us was that we really looked at this is honestly born right around when Opus came out, Opus four six that started really kicking into high gear the ability to trust complex coding systems to these ai models. And that's where the the big vision for this thing really really came from. And I can talk about architecture and like how how we do and don't use AI in the actual collection and running of this stuff. And I hopefully I won't lose too many people, but for the geeks out there in the crowd it might get might get kind of interesting. Lindsey DeFalco: Yeah, excellent. Yeah. And I think like, you know, before we get into the we did we did the why, the what, the why, the why now, and I think the how is next. another thing that I think is unique about just the asset in our crossbeam hands is that you know, anyone who has this data asset and they're running it alongside cross beam but not within crossbeam, they still have the matching problem and like the account mapping problem. Bob Moore: Yeah, yeah, that's true. Lindsey DeFalco: And that's something that has been interesting to us as you know, something that we can uniquely provide is like this data because it can live in cross beam benefits from the fact that we have this, you know, complex matching algorithm and and you actually get the account mapping with this third party data. And so even if you have second party data, third party data data and they're separate, having the ability to actually do account mapping on both simultaneously is a is a really big unlock. Bob Moore: I that's a great point. The you might ask the question, why would you go into a massively commoditized market? where there's already probably a dozen companies selling what at least sounds like or feels like the same exact data through various means, some of which are like, you know, more than a decade old, some of which are many decades old, depending if you include like you know, the the OG kind of data enrichment providers in it. and the answer is because well how do you how do you win first of all there's a market there. Commoditized or not, there's a market. And that market has a certain scale to it. Then the question is how do you enter into a commoditized market and and win? Well, you win on on cost and volume, first of all, so you figure out a way to charge less for this stuff, but still have it be economically rational for you to do that. And if we can manufacture and create our own product from scratch in a vacuum significantly I'm talking like orders of magnitude less expensively than the historical providers, then there's a there's an initial edge, which is we very likely can be as price competitive as the market requires in order for us to start you know, having a interesting or viable business from it. But then the second dimension is is there a way that we can actually offer a differentiated value proposition or a way in which this this data becomes more valuable? And I think there are two dimensions on it. One is by being AI native from day zero. So that it's very, very much kind of a a method of choice or a channel of choice, not just for humans, but for agents who are looking to get this kind of enrichment or other like technographic information in the work that they're doing and making it just dead simple, super easy, and a core part of like the origin story of the business. So there's like an AEO aspect to feel like, hey, the agents know this is like an agentically native place to go and get this stuff. And it's they know it's gonna work. And it's gonna be easy for the agents to sign up, the agents to procure the agents. To use it. That's a big deal. But secondarily, the crossbeam connection, even if we get no customers on trace independently as its own standalone product in like the hyper competitive kind of PLG data enrichment space, it doesn't necessarily matter because every crossbeam customer is also going to benefit from the existence of this trace data without even needing to know that trace exists. Because CrossBeam will be able to lean on all that trace data in the back end to do things like create open data partners, create lookalite. Like populations do embedded enrichment inside of its own product. And because Trace is owned and operated by Crossbeam, then you get this incredible lift that not only increases Crossbeam's differentiation as a product, but does so in a nearly costless manner that just further helps us secure our position in our existing market, where we still have an enormous growth tailwind and a really, really high ceiling in terms of you know, adjustable market ahead of us. So it's that you could look at it as just like. a really cool cross meme strategic thing, but as it turns out, we end up with this asset that on its own is super valuable to all kinds of personas and company types that might never be customers of Crossbeam or might not be for for some time. So we get to kind of have our cake and eat it too. And I think if you were just a new startup in Y Combinator or something trying to build something like this, you too could do it in six months. I don't think there's anything super novel about this. But I think what we have is these this specific connection into Crossbeam gives us this foundational reason why we know it's gonna create many millions of dollars of incremental value in Lyft, even if no one buys it out there in the real And it's gonna save Crossbeam the headache of needing a contract with a third party provider to buy or license this stuff and then have this perpetual financial entitlement that we that we have to deal with because we don't own the data that we're that we're presenting in our product, which is not a position I ever wanted to be in. Lindsey DeFalco: Yeah. Yeah, and I think this is like the whole use case is also one of those interesting ones that has been I don't want to say destroyed, but like condensed by AI. and so I think a lot of like a lot of the use cases for this type of third party data, especially like technographic enrichment, you can just ask Claude or Chat GPT like who you know, what what are the customers Or is this is this account a customer of you know, blah, blah, blah? Or what is their technographic stack? And, you know, depending on how much you want to spend in tokens, it can go off and and do that for you. and so it's like the one-off use cases that I think a lot of people use this data for kind of go out the window for standalone products like this. And then the data asset as a whole and doing this at scale is what gets really interesting. Bob Moore: Mm-hmm. Lindsey DeFalco: and so yeah, that's where I think the you know, again the like having it plugged into Crossbeam regardless of the standalone product gets gets more interesting. Bob Moore: Yeah, I that's a great point. There's like look when you have just a AI prompt, you can say, hey, is Crossbeam a customer of HubSpot? And the AI, depending on the model and its availability to the web, can definitely go out and figure that out. Like it can look at our job descriptions for marketers and see if we mention HubSpot. It can look at like, you know, JavaScript tags or like DNS records or other things to like show clear indications that We have HubSpot landing pages and other things linked from from our website. Like these things are are easy, publicly observable. But if you say to that same LLM, give me a list of 6,000 companies that are HubSpot customers, but only ones that have at least a thousand employees, and remove anybody that was not a customer you know six months ago or something like that. Like I only want people that have been using it for a while, it is absolutely not gonna be able to do that for you because it it requires basically running, first of all, you don't know where to start. Like Lindsey DeFalco: Yep. Bob Moore: Which companies are we even talking about to try and go and collect that information on? And this is where it gets really, really interesting to build this very interesting hybrid of deterministic, classic conventional code and data collectors, and multiply that against the auditing cultivation and rule generation machinery that is best executed by an LL. And that's that's really like the nuts and bolts in the the the heart of how trace is architected. And without going into it too much, just know that if we just ask the LLMs to go out for 10 million companies and go ask them about 20,000 different technologies and said, Hey, does X use Y, we would go so incredibly broke on tokens, let alone keeping that up to date over time. Like it's like this is the thing that this is not is us just asking an LLM in real time for every single one of these data points. That would be absolutely bonkers. But with the yellow intended, but it you could. You absolutely could. If you had an infinite money supply and all the hardware in the world and maybe like a satellite ring of space-based server farms, you could totally go out and and do this, and it would be very dumb and it would be more expensive than hiring a bunch of humans using the old school methods. but you could do it. And I think what what what's novel about this and what's novel about our approach here, and I'm happy to give up Lindsey DeFalco: But you could. Bob Moore: The secret, which is the overwhelming majority of the data collection is still done by very deterministic things that are looking for particular signals that fall under a universe of rules that are like if this particular thing is observed, then we constitute it as evidence and this is true. And it used to be that those kind of rules were cultivated by humans who had to do a lot of maybe they'd write some scripts and they'd kind of like, you know, look for patterns and things using code, but ultimately The validation and verification was a very human task, which required a lot of human hours that had to be done in a not that parallelizable way in order to say, okay, now we have this corpus of 20,000 or 40,000 technologies that we know to look for, and we have these things that indicate that they're in use. And this over time, combined with a bunch of scraping of the entire internet or canvassing of the internet, puts us in a place where We can apply the rules against the data we've observed and we can get this corpus of information out. what we're able to do is take the things that are highly repeatable and can be done very inexpensively, like keeping an eye on the whole internet and you know looking for changes in these various dimensions where these signals might exist. That can all be done very deterministically and inexpensively using conventional. kind of cloud-based methodologies and you know commoditized third party services in their own right that that are able to look for those things and be used. But the core is of rules, like how do we know that we've observed something worthwhile here and you know how do we validate that it is actually accurate an enormous percentage of the time if we use this rule? And what rules are we missing based on the things we're observing and seeing repeatedly in a bunch of different circumstances or dimensions of the data. That can be completely autonomously and perpetually brought to life and kept alive and monitored and cultivated by agentic work and by agentic workers, as can data quality auditing that then goes in after the fact and looks at evidence that we've cultivated and looks for things that are just like obviously flat out wrong. and what's really cool about kind of like this is like a Classic central limit theorem thing from you know, if we've taken an old statistics class, like if you take a random sampling out of 20 million records and you know, you look at a thousand of them and go looking for incorrect facts that Trace is presenting and you identify one, chances are that it's representative of the underlying failure rate of the population. So if we're if we're okay with False positives or false negatives in the data being at like a 0.1% rate or a.01% rate, then we can do random samplings and identify problematic rules or problematic methods of data collection agentically and then chase that down in an autonomous manner that keeps the data set alive and healthy and accurate without needing to have LLMs actually physically look at all, you know, 20 trillion data points if you multiply by all those like yeses and no's against one another. So Our LLM bill is really cheap, but it's extremely targeted LLM use to figure out the like how do we cultivate this rules engine and how do we make sure that the quality downstream maintains a certain level and how do we marry those two? And then the actual coding is done by myself and and a small handful of other engineers that are kind of human watchdogs over this corpus of coding that powers the deterministic stuff. but it is ultimately written by 99% of the lines of code are written by AI and just code reviewed by humans. And then it just gives us this crazy velocity. And that's that's like b because of b having been able to build that, we have this really nice marriage of deterministic and non-deterministic coding and accuracy and data collection methods managed by a small handful of humans. And here we are six months later, and we have this product that I think you know, can play ball in this otherwise highly commoditized market. but with this really specific branch of use cases related to technographic data. Lindsey DeFalco: Yeah, I I love that you're talking about two different things, which is like one, the what AI is actually in the product, and what you know, what we're using the LLMs for to actually deliver value in the product and like what it's actually doing with within trace. And then number two is how you're using AI to build the product. and I think those two things get conflated often and I Bob Moore: Mm-hmm. Lindsey DeFalco: I wonder if you surveyed a bunch of people on what AI native means for a company, is it does the product have to not be deterministic? Like does it have to be, you know, using AI to generate some sort of outputs, or is it just that the entire product is built by AI in terms of like both product decisions, coding decisions, writing actual code? I always like to make that distinction because I think it's I think it's an interesting one. And I'm curious, like, For you, so you guys, you know, built this in six months. First question: how long do you think this would have taken pre pre AI? Bob Moore: Well, I'll tell you because I I built a version of this back at do you have your bingo card out? Back at RJ Metrics back in this would have been around twenty fourteen or twenty fifteen. We had this system called Olga O-L-G-A. It was called the online lead generation automator. And it literally did this. like we ran Olga and we were at the time we had some like co-located server farms, like we were doing our own hosting and we were able to find Lindsey DeFalco: Check. Bob Moore: like dormant CPU cycles and we had this thing that would run on machines that were under underutilized just to get the absolute because we were paying a fixed rate for electricity and server capacity. So anything with empty hard drive space or empty processing, we're always trying to just like, you know, squeeze everything we could out of it. So Olga would run and it would do something very similar to what Trace does now. So I've actually gone and tried to build a system like this before. It's always been like a fun curiosity thing for me. And I probably spent two and a half years building it and built a data set that was one-tenth the size of traces and and like kind of s it it it was useful for RJ metrics. It was actually powering our whole SDR team, like the whole data intelligence for all of them. But it was very narrow. like it was really specific to e-commerce companies, which was our target market, and the data would go out of date, and the signals and alerts were kind of weak, and it did a bunch of stuff very in a very shoddy way. But because it's just like A giant pile of data. If like 80% of it was right 80% of the time and it was covering, you know, 50% of our target market, you multiply all those numbers against each other in a market with a couple hundred thousand prospects in it, and it's like, holy crap, we have a thousand really good leads here. and it worked. So I believe that what's interesting about this kind of business is that in order to make it work and work well over long periods of time, you have to run on a treadmill. So there's building out the initial thing, which I think is a years long proposition for a small team of really good engineers, or maybe shorter for less talented engineers that might be less expensive, kind of all thrown at the problem. but even if you get that done, you still have to keep it up to date. And not only does the data change, like literally companies move from HubSpot to Salesforce or vice versa, and these are things you need to track, and and the system needs to keep track of that. But the rules themselves change because companies get acquired. you know, new technologies emerge, new ways of doing business emerge, new s old signals go bad and new signals go good. So your rules engine and your rules database needs to stay up to date, and the companies and vendors you're tracking needs to stay up to date. So even if you build the perfect version today, if you wait six months, your data set's gonna be wildly out of date. So how do you keep that going? And how much of that can you actually productize versus you have to throw humans at conventionally? So I it's not just about how long would it take to build it, which I do think was sped up by some, you know, what I will call a 10x factor for the level of quality and coverage that we have. Sped up by 10x. So six months, you know, turns into five years. but also Lindsey DeFalco: Mm-hmm. Bob Moore: I believe the maintenance burden of keeping it up to date and maintaining really high quality is also at least an order of magnitude better. not to mention there's less stuff that's gonna be out of date for for less time, right? and that's the other thing. You can like inside the product, you can flag something as being inaccurate or like suspicious, and immediately like it kicks off one of these like LLM data quality auditors that goes and like. Lindsey DeFalco: Mm-hmm. Bob Moore: First of all, it's like a little skeptical, like, are you spamming us? Do you have an ulterior motive? Whatever. but then goes in and and cross references the thing that you claim is wrong and looks at the real evidence and like either confirms that we had it right and the user is wrong, or it goes and actually like takes corrective measures to like reshape not just the data that was flagged, but what's the underpinning reason why that data was identified? Do we have a bad rule? Is a piece of evidence we were relying on no longer there or no longer accurate? and it will go and evolve dynamically. So human inputs also help shape you know the fabric of how this all gets gets pulled together too which is which is pretty cool. Lindsey DeFalco: Yeah, and I I like that the like the intentional application of when to use AI versus when to do something deterministic. And I I think we did talk about this before, but like this is such a big thing with agents right now. Like everyone has a different definition for what an agent is and a lot of agents are still you know, I'd say like eighty percent deterministic. And like that's okay. And it I think it's because a lot of processes and things that people do do have a certain path that they follow or need to follow. So, you know, if this then that and like you can't make stuff up in the meantime for various reasons. and I think we we talked a little bit about this with with internally with with matching as well. It's like you don't want it there's some hard rules that you want to be followed always. But at the edges, like with what you're talking about, if there are questions or issues, Bob Moore: Yeah. Yeah. Lindsey DeFalco: That that use case where you used to need to have a human come in and reason about it. Like that's where we get back to like AI is very good at the reasoning. and so inserting it in the places where it can make the deterministic outputs better, but not replace the determinism is it's I think where like the highest, highest leverage use of AI is right now. Bob Moore: Yeah. Yeah, I think that's really well put. it it's it's it's this marriage of how do we continue to do the commodity stuff in a commodity way, for cost purposes or frankly speed as well. Like it's just way faster to canvas the internet and look at a million websites with deterministic code than it is to rely on an L L Lindsey DeFalco: Yeah. Yep. Bob Moore: kind of sandwiched on the input and output side of doing that. So not only is it significantly more costly to do it with an LLM, it's just like it's slow. because there's a lot of compute that that needs to be the bread of that sandwich that is completely unnecessary and takes chip time and and and dollars as well. Lindsey DeFalco: Mm-hmm. Bob Moore: love I I wanna I wanna show one thing off from from the trace product because you because you brought this up, I actually think it's a really it's a really good point. Like where we do and don't use AI inside of this thing, again, trace can be totally interacted with through MCP or you can talk to it through You know, through your favorite through Claude or through ChatGPT or Gemini or whatever. You just connect the MCP server. So you can use this in a completely headless fashion. However, I'm gonna log into the actual UI here because I want to show you a great example of I love this pattern, this UX pattern that exists in this moment today where there are there are human users still using these things that are are very agentically compatible. So in trace you can generate these things called lists. And much like in Crossbeam you can generate a list and a list is just like, hey, I want to see everything that overlaps with my partner or hey I want to see everything that you know has these properties. It's basically a saved filter. Okay, so let me pull this up and show you where we're just relying on an LLM or relying on chat here to just make this significantly easier than a traditional builder. I can say show me companies who have hubs who use HubSpot and have over 1,000 employees. And what it's going to do is it can actually go in and it knows very logically how to parse that out into a rule system that fits directly into the list generation engine behind the scenes. save as a list. And what it's going to do now, and this is the pattern that I love, is it actually goes and takes you to the list, right? So the the chat that lives in Lindsey DeFalco: Yeah. Bob Moore: The product will navigate you within the app to the place that you wanted to go, right? And now we have this list. I can save it. I can get notified via the app or email or webhooks when there's new things that get added or there's things that change about those. I can download it as a CSV and it'll, you know, tell me how many credits it's gonna cost me in order to go and export all this data, which is kind of the the financial model here. And all of this, so there's 5,600 companies that that's true for. And you know, we can go into any one of these and just see more about like if we want to go look at Ugov, we can see the technologies they use, how many partners they have, how many jobs they're hiring for, right? And all this information is is super interesting to to look at. So I think that is that is this very interesting thing where you're able to bring in this synthesis of the jumping off point of just being able to use natural language, but then it's physically moving you. You don't even have to click inside the app. It's taking you to the UI component that it has created as a result of your chat. And man, I love that paradigm. And I think that's something that's really we want to do that in Crossbeam. And I think we will do that more and more, but it's a lot harder when you start with a really robust complicated UI and then you try to layer chat on top of it. But when you're building in this AI native way where you think about that directionally as the usage pattern, it becomes So easy to just have it do it and take you there in a way that's always going to work. And I I I love that pattern. It was really fun to build that. Lindsey DeFalco: Yeah, this is where the AI starts to feel like like Jarvis to me. it's like when you're you're mixing like cause I think we were all so blown away by you know, how smart that's arguable. you know, AI is in terms of being able to reason and have conversations and access information. but I think like I very quickly hit a wall where I was like, Okay, but I need you to like do things. And then we got like cowork and we got, you know, we're Bob Moore: Mm-hmm. Lindsey DeFalco: We're we're there, we're doing it. but yeah, in terms of like building new products, having that same experience that we have in all of these AI platforms mimicked, where the things that you don't feel like you should need to select dropdowns and, you know, scroll through different options and things like that, having that be married with. Bob Moore: Yeah. Lindsey DeFalco: the like all of the magical feeling benefits of these these you know lists and things just getting created for you. yeah, I I I love that as a UI direction. and I think it's one thing I found the more I'm using, you know, like Chat GPT, Claude, to ask questions in my daily life is I go to Google still sometimes. And it's funny 'cause you were talking about like speed being a a limiting factor in some of this stuff. And like sometimes I'm like, I don't wanna sit there and watch it think and have it come to the wrong conclusion when I just need to look up like how many, you know, ounces are in a cup, which I still don't have memorized. so it's funny too, it's like where Bob Moore: Yeah, who voices the bad guy in Hotel Transylvania? Like I'm still going to Google for that 'cause I don't want to wait like forty five seconds. It's Jim Gaffigan, by the Lindsey DeFalco: Yeah, yeah. Exactly. Yeah, looking up like Yeah, the old cast cast lookup for every movie I'm watching and what else they're in. yes, what's funny because I think like w we're still in the middle of ourselves learning what things we have patience for AI and is the result better enough for us to actually wait for it, or what things we just kinda wanna like look up. which is Bob Moore: Yeah. Mm-hmm. Lindsey DeFalco: Yeah, which goes into how you're actually building this thing, which is interesting. Bob Moore: It kind of mimics human the human brain and human knowledge as well, which is like there are things that we commit to memory, kind of our you know, system one and system two, like automatic permanent memory versus short term contextual what I'm working on today memory, that becomes very rapidly retrievable. Lindsey DeFalco: Yeah. Bob Moore: versus stuff that is kind of buried deep in our knowledge and we have to think really hard about, andor you know that you know somebody who knows and you're willing to pick up a phone or find that person and you kind of know this because of the knowledge graph of your social circle. it's as you look at AI as something that mimics evolutionary tendencies, it also there's a there's a very interesting set of analogies there. that, you know, pick pick how creeped out you want to be by it. But it it kind of makes sense. it's like RAM versus hard drives versus cold storage versus cloud. it's all you have all the options available to you, yeah. Lindsey DeFalco: Yeah. Getting real nerdy now. Yeah. Yeah. So I cause I think that we just talked a lot about like the in-product, how to use AI and the different decisions you made in terms of, you know, the UI, but also just how to handle the data quality and data maintenance. is there more around like the how you have built this thing using AI that you want to touch on? Like how small the team is, what the process looks like. and how different it is from classic classic building when it's AI native. Bob Moore: I'd say, you know what's been fun about it is watching the models get better just in these six months. I would say that when we started out, the models by far were the weakest on infrastructure. So and and this is a pretty infrastructure heavy product, right? So you know, we're hosting it in one of the hyperscalers and you have to, you know, run A lot of very discrete things from the web hosting to the various means of storing and retrieving the information to the actual things that do the tracing that go out to the web to retrieve this info, to all of the security and hardening that needs to happen. And you know, you can use use tools like Terraform to kind of stand a lot of that up. And you can do infrastructure as code, which is kind of the method, but I had done a lot of that. That is not my I am like an algorithm. algorithms and data structures guy and a databases guy and I'm not a necessarily like a DevOps expert and I had gotten pretty far with Opus doing a lot of Terraform work and had built something functional. And then our CTO Alex, you know, I I got him plugged in on this project a month or two in and he came in and just completely gutted all of the work that had been done on the infrastructure side and to great effect. Like saved a bunch of money was better organized. It was easier for new engineers to come in. Like Lindsey DeFalco: That's it. Bob Moore: the the LLMs were just not that good. And now by now we've gotten to a place where we're a couple of subversions of Opus later and occasionally you can point fable at at what you need if you can bear the the credit cost. And it's like pretty darn good. Like it's it's it's gotten it's gotten to a place. And I think part of that maybe Alex's work then helps set a tone for the direction that it needs to go in and then it needs to mimic when it is doing work. But I think also inherently the models themselves are able to like put on a DevOps or system architect hat in a much more realistic, reliable way. so that that's just one example of many of how it has gotten better. The UIUX stuff, it's been very clear how that has gotten better, including the tooling around things like design prototypes and you know, what we can do. soup tenant like Taisha, who's working on the UIUX now, Who is one of our in-house traditionally a a brand designer who has kind of brilliantly expanded her her scope and mandate into a lot of these AI assisted like UI UX and product design chops, which is is so fun to watch. doing a great job and moving really, really fast. And it's it's through the aid of this stuff, but her design eye is still better than what the AI can create. But the tooling for someone like her who is not an engineer. by trade or or even an HTML programmer necessarily or CSS programmer. had some chops there, but like it's not the like she's a designer. is able to now be extremely full stack extremely quickly to great effect. and and I think that's been a really cool thing. It's like the opportunity to use this as a platform to bring people along on that journey and provide access to new ways for people to cultivate and and flex their muscles and build out skills on that has been really fun. So yeah, I I think all that you know there's there's There's a bunch of of learnings and observations from it, but it's all coming at us so fast that I think it's it's kind of a blur. We're we're in it right now. Lindsey DeFalco: Yeah, it's the the thing that stands out to me about all of that is that you're really talking about AI accelerating people's natural skills, which to me has been the most fun thing to watch about this. And on on our our product and design team we call any AI design, we call it like Tamu design. Because it gets like most of the way there. It's like pretty good, right? Like and you know, if you're not looking too closely, it's an imitation of the of the real thing. But, you know, our product designers still have the taste and the skill and experience to to know what a good final experience, you know, visually and UX wise is. And so I think, you know what you're saying about like Taisha having the taste and the experience from a design perspective and then Alex having all of his years of you know, background and skills on, you know, infrastructure related things, like and I think you and me actually talked at one point in one of our one-on-ones about like you have a very strong product mind and doing what you're doing to lead a project like this, and I'm not trying to boost your ego here, but unfortunately it's gonna be it's gonna be a compliment. but like not not everyone can get things as far as you are able to because you have a very good product mind and like your ability to zoom out and look at the strategy and then zoom in and look at the the technical is very strong. And you know, we talked about that when we're talking about Bob Moore: no, go on. I'll brace myself. Lindsey DeFalco: really bringing bringing teams along and what really strong profiles look like in these AI native companies. And more and more it's that blend of some of the skill sets we just described. It's like, do you have design taste and technical knowledge? Do you have product knowledge and technical knowledge? And like you have to be somewhere in the Venn diagram of the the the different functions to be incredibly valuable. in this era and so it's it it's been really cool watching watching Trace get get built 'cause it it mirrors early product building in a lot of ways, but then it's it's so accelerated because of, you know, all of the things we just talked about. it's it's really cool. Bob Moore: Yeah, I think there's something true about startup hiring and startup teams in general, at least when done well, that is so well suited for this era. Because I I've always had kind of a a cultural default to there's no such thing as not my job, inside of a company that is, you know, in the in the hundreds of people as opposed to the thousands or ten thousands, and you become forced to kind Lindsey DeFalco: Yeah. Bob Moore: you know, you go from what is that saying? It's like Pirates versus the Navy, right? it's it's some metaphor like that. Like as you get larger, you have to specialize, you have to compartmentalize, you have to organize, you have to have these kind of like regiments that exist so that you can continue to function and know what's going on. So you become this very mechanized apparatus as a very, very large company. And what it what a byproduct of that is that people become hyper-specialized in what they do and the scopes of work become more narrow. And this is where you get a lot of this like not my job isms that kind of start showing up because. 'cause people stay within their lanes, and it becomes kind of weird for somebody to, you know, encroach in other territory. Whereas in a startup, you have to, you know, Do like cabinet making some days, and you have to go do design other days and have to build stuff other days, and everybody has to know how to talk to customers. And that's kind of the nature of the beast. So when you get into a world where AI provides an opportunity for people to get a lot more leverage out of their time, it means they can do more, and it means that a willingness or an interest in doing more, in zooming in and zooming out, in taking on the adjacencies to your traditional role are like really, really enabling and powerful. But This is why like we we haven't laid anybody off for AI. And my my goal and hope and expectation is that we never have to because what's happening is we're growing the business at a really considerably high percentage and we're not hiring much at all. And we are able to lean into the fact that we've built up a company that's made up of people that are pre-vetted to not have a problem stepping outside the lane of what they might think their job is and actually be expansive in terms of what they're able to contribute and what areas they're they're able to work in. And we're seeing that work in so many really cool ways and kind of finding the right the right structure to put all that into. But I think when I see a 10,000 or 100,000 person company lay off thirty percent of their staff, I d hate that it's happening for those people affected, but I get it because a lot of those companies have organized themselves and hired in such a way where there's this narrow scoping that exists around a lot of people's job responsibilities that actually make them very well suited to be replaced by AI and make the people in those seats not extremely well suited or predisposed to adapting or changing. And I I think in those larger companies it becomes more more inevitable. Whereas in a startup if you've hired well and built a robust team that kind of showed up willing to make cabinets some days, I think you as long as you have good product market fit and you actually have a a product that the world wants, then AI can become a tool to have a really awesome growth rate and just create opportunities for the people in in inside your doors. And look in a lot of companies it's a hybrid and and two things can be true at the same time. But my aspiration for Crossbeam is that we're we're very, very much in that second category. And that's that's what's been playing out so far, at least in the last year of being in kind of this post-late stage AI world. Lindsey DeFalco: Yeah. And I think like, you know, what goes along with making people, especially people who are really good at wearing lots of different hats more efficient, is that you need less of them, which is what plays out in the bigger companies, which is exactly what you're talking about. And then in smaller startups where, you know, intentionally have startups that have intentionally stayed lean, it just means that you're getting more efficient. and yeah, it's been It it's been fun for me watching you know, when we're getting past the hype cycle of like, I AI can do everything. so like you were saying, you're like, yeah, I'm gonna be like the ops person for trace and then hitting that wall and being like, wait, no, I shouldn't be the ops person for trace. Like those walls are getting hit everywhere, which I think will dictate the shape of of organizations. and it may mean that, you know, you need someone to Bob Moore: Yeah. Yeah. Lindsey DeFalco: you know, to not be like full time on it, but you need someone part time. but I think that's happening everywhere, right? Like, you know, you're looking at prototypes from a design perspective and you're like, Yeah, this is this is so crazy it can make this. but then you're like, Well, also it kinda sucks. and so we're starting to hit those like the limits of what what AI should do and what you still want the humans that have the actual taste to be the final say in. Bob Moore: Yeah, and it's a moving target. and I'll say it the one thing I'll say about Trace that i i is kind of a interesting truism is like we're about to launch this thing and go to market with it. And Lindsey DeFalco: Exactly, yeah. Bob Moore: How hard it is to launch a new product in this era is something I'm terrified by. Because the noise is astronomically large. There's so much noise. no one has any attention span. Everyone has enormous expectations. and I'm not entirely clear on the ways in which you can effectively in a full stack world. Lindsey DeFalco: Hm. Mm-hmm. Bob Moore: turn AI and AI systems into something that actually works and converts. it kind of is the the belief I've always had about SDR teams and and paid ads and all kinds of other stuff is like if you have great product market fit, you look like a genius and you look like you really know how to run the best ads in the world and create the best SDR program. But the reality is like there's a latent demand in the market where discovery and awareness is the only missing step. And I think that's kind of true probably with Trace, but because it's a commoditized market, like awareness that solutions like this exist is maybe like not all that big of a aha moment for people. so I I'm struggling to think of ways to deploy AI against that go to market process without Lindsey DeFalco: One last. Bob Moore: leaning on something that's gonna actually still require a lot of conventional human work, and it'll be kind of a human hybrid at least initially. So I think in the first few months of this thing being out of market, we're gonna learn a lot. about how far AI can take us on go to market. And I think it's a lot it's a lot less far along and there's a lot less well established, like, sustainable best practices on that compared to the engineering product design and and product management pieces of of what we've been able to build. which is the curse of everybody's like, this is no vibe coded product, but like, you know, zoom out and squint. It's kind of a vibe. It's like it's it it it looks and smells like something that was vibe coded by some some crazy founder and the rubber meets the row when these things try to go to market and they can't get anybody's attention. and I think that's gonna be the big test that's coming up is like, okay, we spent six months building this data corpus and making it great, now does anybody give a crap? and I think that is you know, we know Crossbeam gives a crap and we're gonna create leverage for it, but how do you actually sell something in today's world that's starting from nothing? That's gonna be a big, a big lesson that's about to hit us. Lindsey DeFalco: Mm-hmm. And when you're competing even more and more with people just building things in-house, and when you have podcasts coming out saying how easy it is to build an entire product like this in six months. Yeah. Crazy. no, but I mean it's true. It's like the the balance of people are like, hey, with AI, I can just build this in-house. but then also on the flip side, I feel like people have Bob Moore: Yeah. And like giving the whole playbook away. Lindsey DeFalco: more patience with their vendors than ever because they trust that their vendors can build things more quickly with AI. And so there's just like weird balance between like, I think in some ways it's creating more lock-in for existing vendors. but then also a higher risk of being, you know, rebuilt in-house. and so who knows how that's gonna play out. But I think that's a Trace is gonna hit a sweet spot there where it's like Bob Moore: Yeah, yeah, yeah. Lindsey DeFalco: I don't know, maybe we can get close enough. and yeah, your point of like actually getting people's attention. I don't know. Good luck. Bob Moore: Yeah. or maybe we're racing against maybe we don't need people at all and what we need to do is get the agents' attention and have them just think that like this is the best data and it's the easiest for them to interact with and and maybe it's the least expensive as well. and that's actually the solve, which I I think is kind of a fun Lindsey DeFalco: Yeah. Bob Moore: a fun direction to go in. like allowing agents to actually register for accounts from scratch is one of the next things that I'm I'm looking into right now because there are you and I have talked about this like there are some some standards that are showing up out there for agents to associate themselves with a true identity of being representative of a true company that they're working for. we may build a few experimental things on the backs of that and see see what we get. Lindsey DeFalco: Yeah. I feel like that was every my god, I just forgot the name of that company that like blew up overnight and got acquired by OpenAI Clawed OpenClaw, yeah. It was like every example of OpenClaw was the agent being like, I need technographic information. Trace provides that. Let me go get a trace account. Okay, I now have access to technographic information. but you know, not B2B. It was more people trying to check in for airline flights and things like that. But Bob Moore: Open force. Lindsey DeFalco: Yeah, it's exactly that. It's like maybe everything you just described actually doesn't matter because it's it's not the humans making the decision. Then we get back to eight. Yeah. Bob Moore: Yeah. Or it's the humans asking the AI to go do it and they go do exactly what you just described and hand over the credentials. yeah, kinda wild. Well, hey, this is great. I'm glad we got to talk about trace a bit, and yeah, look, if you're a listener to the podcast and you want to kick the tires, it is free to start. You get a bunch of credits just for signing up. at least for now. So go tracedata.ai is the URL and I'm easy to find. Let me know what you think. Hit me up on LinkedIn or Lindsey DeFalco: Yeah. Yeah. Bob Moore: You can't. Lindsey DeFalco: It'll also soon be listed in Chat GPT and cloud marketplaces. So you can just Bob Moore: that's true. that's gonna be wild. Yeah, we've got a bunch of partnerships brewing up that may show up by the time this this goes live. Lindsey DeFalco: Sweet. yeah, this was a good one. It was good to hear about what you have been working on over there. what what would you say you do? Bob Moore: Cool. What exactly would you say it is that you do here? All right. It's always a pleasure. See you all later. Lindsey DeFalco: By