122. Timing the AI Wave with Brian Raymond | All Quiet on the Second Front Podcast

Speaker A: Foreign keeps saying we're at, uh, this inflection point. I've been skeptical of that because the models have been really good at some things, but then there's actually, there's been, like, a real delivery gap between, like, what you can demonstrate in a POC or you can, like, vibe out of it, but what you can actually put in hands in production. Uh, and I feel like with four seven, with Mythos coming, with Cloud five, five that just dropped with others, we're now at a point where that 95% failure rate is going to start declining pretty rapidly.

Speaker B: This is All Quiet on the Second Front podcast where boring conversations around defense tech and national security come to die. Join me, Tyler Sweat, and my second front homies as we dismantle the mundane and cut through the bureaucratic bullshit to demystify the world of defense tech. But be warned, this is not a typical government podcast. Are you ready to get weird? What's up, everybody? Welcome. I'm your host, Tyler Sweat. Welcome to another episode of All Quiet on the Second Front, the podcast where boring defense talk comes to die. We're gonna, uh, we're gonna do the unthinkable. We're gonna run one of the OGs back a little bit. So I, uh, think one of one of my favorite folks, longtime friend, second time caller now, Brian, and the unstructured team. Um, brother, thanks for making time to come on back and, uh, and do a little update.

Speaker A: I'm delighted to be here, man. This is exciting. Thanks for having me.

Speaker B: Hell, yeah. So, I mean, you and I, you and I have known each other for, I mean, as long as I can remember at this point. So I'll do a short little, hey, hit him with kind of who you are, what you're working on, and then I want to get into the. Hey, cool. We chatted years ago about how we were thinking about early stage corporate leadership and what's that mean? And I'd love to bring that into the contemporary sort of world in reality and say, all right, what's changed? What hasn't? All of that.

Speaker A: That sounds great. Well, maybe I'll start off by just sharing a little bit about my background. And it's a background that's not too dissimilar from a lot of folks in defense tech these days, in that I've had a lot of zigs and zags. So I cut my teeth at CIA. I spent a bunch of years over at CIA as an analyst, spent a decent amount of time in the field, and then moved from there down to the White House. I was country director for Iraq during all the ISIS stuff and then did a hard left turn, um, career wise. Um, this is about 10 years ago. So did an MBA, landed in investment banking, didn't enjoy moving logos around slides and doing dcfs and was fortunate enough to find my way to a really cool um, startup at the time. This is around 2018. So right after transformer based models were introduced called Primer AI. So they had just raised their Series A and I got a four year education in natural language processing and ML engineering. There was fortunate enough to be able to lead um, a chunk of their business in the natsec so space. And then in 2022 a little less than four years ago, I broke out of my own, raised an initial round for unstructured. Uh, to date we've raised about $65 million. We're about 80 people on the team and we focus on problems kind of to the left. So what we do is we take any customer data and get it ready for gen. The least sexy pitch for it is we'll take your PDFs and PowerPoints and videos and images and get them into chunked embedded JSON. But we do a whole lot more than that. And so that's, that's me in 30 seconds.

Speaker B: Yeah, dude. And so you know, we talk about 22 kind of making the jump, right? Like the world has, has changed. There's been a few rotations since then. I think we've seen some changes geopolitically, I think we've seen some changes in sort of capital market. I think we've seen some changes sort of in uh, the broader technological sort of reality. Right. Like what, what are sort of, you know, table stakes from a technology standpoint in an organization versus what are maybe first movers, um, sort of thinking across those paradigms, like as you look back now, what do you see that like maybe you got right or you got wrong as sort of how you looked out on the horizon from that point.

Speaker A: Look, one thing that was impossible to predict at the time was how quickly models, larger models would progress in their capabilities and both in terms of anticipating and progressing too slowly because no one really saw the chat GPT moment coming in November of 2022, but then also how quickly they'd improve. And so there have been a lot of false starts as well. And so during this period of time, like it's still like still today in 2026, it's still difficult to effectively parse a table and a PDF, which sounds silly, right? On the other hand, we're worried about Mythos and AGI getting out and hacking into our government systems. Right. And so like, uh, we're in this weird reality right now. And so building a company around this tooling set and around this industry has been one in which timing matters a lot. The timing in which you're racing hard and you're moving very quickly, but then also maybe when you step back and you let it play out a little bit so you can gather a little some more data, a little more confidence in some of these key bets, that has been the most challenging thing over the last three years is just figuring out like what's real, uh, like what's reality and what's not reality with what's going on in the space.

Speaker B: Yeah. And I mean, I think share your sentiment there on understanding, hey, you know, what's hype? What do you push chips on? You know, what do you push chips in against? And what are you not, um, what have you seen thinking about that sort of like evolving, you know, perspectives around reality? How have you seen that either be consistent or maybe divergent across like the capital allocators, the, you know, government or sort of highly regulated industries and then you at the sort of like technology corporate operator. Have you seen, you know, as you say, hey, this is real. Are you then trying to convince a VC who thinks this other thing is real that you guys are on different pages or what's that been like?

Speaker A: Look, it's, it's um. I think we're always reacting to things like first principles, right. So what are we seeing? First we saw Nvidia go absolutely bananas at the hardware stage. Right. And then we saw the model providers which were feeding a lot of this. Right. So OpenAI and then anthropic really pop. And then you saw some vertical solutions or more packaged solutions really pop. Right? So like glean, cursor, lovable, et cetera, set records for how quickly they hit, you know, 100, $200 million in ARR. And so you see that going on. But then on the other hand, you have organizations that maybe are in regulated industries that have these enormous capex budgets for it, but they're extremely leery or they don't have access to the GPUs, or they're seeing the performance of these things, saying these things are hallucinating like crazy or all of people. We built this chat bot, you know, for our, uh, wealth management teams and they're just using it for their fantasy football leagues and they're not actually using it to drive value. And so, um, there's this, there's, you know, and that I think speaks to this perceived value gap, right where there's a lot of companies that were able to monetize process and not output. I think that one of the first companies that was supposed to start monetizing output has been Sierra and they've been successful. They've raced towards like 100 million ARR, but not a whole lot others. You've got a lot of ink spills around outcome based pricing but still today you don't see a lot of folks actually able to put that into practice. Like they love talking about it in conference panels but like you don't actually see a lot of it out in the wild yet. So that's like this ecosystem, right that we're operating in that we're trying to feel our way around in, if that makes sense.

Speaker B: Does make sense. And it brings me to sort of the next point which is okay, cool if there's this much um, the right word, uh, tension between perspective or perceptions and reality and then there's the sort of change geopolitically, there's markets moving, there's how has it changed how you approach sort of running a company, building a team, you know, thinking about, I know I struggle with like hey, like do we make this higher? Do we not need this higher in 6 months is this thing that gets obe by the next release? Like how do you think about that?

Speaker A: Well, let me get real, I'll just tell you some things I got wrong here and um, and some things that we hope we're doing right now. But one thing that we got wrong, um, or that we did, I mistimed I should say this and, and it's. We didn't build ourselves into a cul de sac but the timing was off, was in 2023 and 2024 in my world, um, deep in the gen AI infrastack from a capital allocators perspective you saw lots of vector DB companies emerging and raising monster piles of cash at ah, enormous valuations on the expectation that one or more than one of them would displace Snowflake or databricks or other folks that built their chopstick, um, during the modern data stack era. What happened? Industry decided we're not going to go pick up all of our data, embed it and store it in a vector database. We're not going to do that. We're going to sit on the sidelines, we're going to continue to prototype at a 95% failure rate or buy the boxes. The gleans that we know, work and then wait for the tech to mature. And so from our standpoint, we built a heavy duty platform to go feed enormous amounts of data to vector databases and then nobody was moving that data. There was instances where they were right. But from a industry timing standpoint, it's like it did it now. Um, we had this thing built early last year, mid, um, last year, and we're like, okay, this is great, we're ready to rock and roll. But then there, there's actually not a lot of data needing to be moved yet because folks are sitting on the sidelines still waiting for agents to get better.

Speaker B: Yep. Then.

Speaker A: And like then in late last year, models got a lot better. Right. Um, you said GPT5, you had Claude 45, Claude 46, now Claude 47. That are beasts. And the agentic architectures could be a lot more successful. Data starting to move. But we were about a year early, right. In terms of our tech timing. So that's like product strategy perspective. That's tough. That's really, really tough. On the company building standpoint, we saw cursor going bananas just like everybody did the about, you know, the winter like 14, 15 months ago. Right. And so we're like, okay, we got to, we got to be cutting more PRs, got to be building faster with the headcount, we got to be more efficient. Right. We're hearing this from investors, we're hearing it from the market, and we're like, okay, let's go do it. And we realized two things. One, um, first, out of the gate, our code base wasn't ready and our documentation wasn't ready to actually use these things. And then two, we didn't have sufficiently senior engineers to have the systems level thinking to be able to actually use these things. And so we spent the spring and summer doing that. But it wasn't until those new better models came out in the fall that things really started to show up in the data in terms of productivity. And so this is manifesting in lots of different areas, um, across the business today.

Speaker B: Yeah, it's interesting to hear you talk about sort of the, this senior sort of like engineering mindset, the ability to apply sort of systems thinking to an ever growing litany of like tactical tooling. And to be that sort of like orchestration maybe is the right word. I think there's, there's and we get it all the time. And I've, I'm going to steal a lot of your language on that because I don't think I've done, I've been as articulate as, hey, like, I can give, you know, uh, the Russian space monkey. I can give it like Claude code and like it will shit out in app. Yeah. But like, is that app gonna function in the next 10 places I needed to run? Is it going to have like the extensibility, the durability and that trade off is real. And we had to do the same thing and leveled up our engineering leadership in like a really, really material way. And it's transformed the volume, velocity and quality. All three have to hit all three of those. That equilibrium we can hit at scale now.

Speaker A: Yeah, yeah. And it's, I think that, and I guess it's, it's this way in probably every big technology shift. But you're scrolling Twitter, you're scrolling LinkedIn, you're talking to folks and you hear these stories from like anthropic product managers, shipping product. Oh, we built cloud code in an afternoon. We just vibe coded it. And then a lot of this is coming from the model building community. But when you're in the business of selling enterprise software, those three pillars that you articulated are absolutely must because you're in expeditionary environments, right. You're in regulated industries. And you're not selling a model, you're not selling an endpoint, right. You're selling an integrated solution. And so, um, it definitely, it's a good competitive pressure, right, to have. But also the reality versus like what you see online mismatch is actually pretty, pretty stark.

Speaker B: Yeah. Where have you, um, how has it been like, employee wise and thinking about like, you know, the, the purists who are like, no, no, no, no, like I write code, I don't have a thing write code. And you know, the folks that are maybe more fully sort of like AI native and are like, yeah, dude, whatever. I talk to mine and it just codes for me. And then I sort of orchestrate. Like, how have you seen that from a.

Speaker A: Uh.

Speaker B: Because I know for us, like, I didn't always hire someone who's like going to be comfortable with like whatever new version of a capability came out. I hired someone that was good at the thing that we needed to do and I got a bunch of that wrong. How has that been for you guys? Especially when you talk about sort of like shooting, shooting the gap, like 12 months left maybe of like when industry caught up. Like, what does that tension look like? And how do you as the CEO, sort of try to harmonize that?

Speaker A: I break it into two pieces. I'm like, okay, how do we upskill or change patterns and change expectations with who's in the door right now? And then two folks that we're bringing in. How do we make sure that they have the ethos right and the hustle and, like, the mentality and the. And the skill and the tooling that we require? And so on the first part, it was like, there's some stuff that you just. The fixed investments that I mentioned earlier that you got to make, you got to the documentation, the code base updates, et cetera. But then after that, I mean, we've had a lot of success using a tool called dx. Um, DX has become extremely popular. It was very uncomfortable in the beginning because engineers and engineering managers historically hate measuring PRs, throughput and other metrics. But the tooling's evolved a lot, and I think that it's created some, uh, it's held up a mirror to folks in constructive ways to say, look, it's time to evolve and we need to adopt new patterns. And, um. But we've had to be sensitive to how we've done this so that we can keep our best folks, but also nudge them along in the direction on the interviewing side. Look, we just say, show us your setup. How many tabs are you running with Claude code? Um, how are you monitoring it, what you're like? And we just say, we don't care what you're using. If it's Claude, if it's cursor, if it's some other setup, we couldn't care less. We're just, um. We just interrogate their rigs effectively and say, we don't care what you build. What we do, though, is we also give them, um, a portion of our code base. And we say, we want you to build, um, this new feature. Here's the pr. Build it. Live with us in the next two hours.

Speaker B: Oh, dope. I like that.

Speaker A: That's been in place for about six weeks now, and I think we've hired three or four people that way. And so it's a. It's a little awkward. We're figuring it out, but, like, how else we're going to figure out if these folks are going to be able to actually capitalize on the tooling, you

Speaker B: know, I know, um, you know, as you look out over the next 12 months, what's got you. What's got you excited?

Speaker A: Yeah, we're, uh. Everyone keeps saying we're at this inflection point. I've been skeptical of that because the models have been really good at some things, but then there's actually. There's been, like, a real Delivery gap between what you can demonstrate in a POC or you can vibe out of it, but what you can actually put in hands in production. Uh, and I feel like with 4 7, with mythos coming with cloud 5.5, that just dropped with others, we're now at a point where that 95% failure rate is going to start declining pretty rapidly.

Speaker B: Yeah.

Speaker A: And you're going to start seeing proliferation. Like just me personally, I just, I almost work almost entirely all day from Claude Cowork that, that, uh, I bas and that's my tool of choice, but I just work there and it might like I'm able to do so much more through this. And so I feel like almost like a dam has broken here over the last quarter or two and that that's going to take years to actually like spread throughout like the economy. Right. But we're at this historic moment where that's occurring and there's lots of opportunities to win here. I think that the trickiness when it comes to being like a VC backed startup is that's like, you know, we were talking about this before the pod kicked off, but about my. In my mind, everything kind of feels like a game of jump rope. Right. And it's like, when do you jump in and how do you time this stuff? Because you're not just like timing the market, you're timing funding rounds. Oh yeah. For growth rates. And how do I map this? And adoption and marketing spend and engineering effort and um, industry, which vertical I'm going to make a bet on. And so we're seeing these just insane J curves all over the place. But, um, it's really tricky to figure out which J curve to make a bet on. Right. And at what point on the J curve to go all in on it, if that makes sense. Yeah.

Speaker B: Are you seeing the J curve? Right. Are you thinking that you're at a certain part of the J curve when you're actually on some other part? Yeah.

Speaker A: And so like, for me, that broader observation that I articulated means like, okay, I'm having more confidence in some of the bets that we're making. Right. And more confidence that it will pull us through. But, um, but it's still dicey, man. It's still. I mean this is, this is the whole. Your space, my space that we're all in as. It's, it's, it's dicey.

Speaker B: Yeah. What? Um, so give me the, give me the other side of that coin. Right. Like you look out over the next 12 months where you're like, ah, fuck, this is going to be a problem.

Speaker A: I think that for us we're developing a bunch of new capabilities and it's like cat like a lot of the new tech that's coming off, we're getting great receptions from companies but for us personally, we get the most market pull from the customers that are among the most challenging uh, to work with. And I'm not saying that like as a badge of honor, like we're going, we did the best. It's sort of like, man, I wish we could find a mid market or SMB segment to scale with because it'd be a lot easier.

Speaker B: Right?

Speaker A: And so for us, we've spent the last year building a ton of tooling to get really fast at NVPC or bare metal deployments. Um, but now it's like, okay, how do we scale across these organizations, provide 24, 7 support, all of these regular scaling things that everyone's like, oh, you just bring the right folks in, you bolt it in. But in reality it's like, okay, this person's calling at 3:00am um, from India, they got a problem or from Japan and they're in this weird instance of Azure and this over here and stuff. Like we had an insurance client and something broke. They perceived that we were on the hook but it was actually some weird Azure single tenant thing on the Azure side. It took two weeks to figure it. It's like that sort of stuff that's like not sexy, not headline stuff, but that is the reality of like scaling like in this gen AI space right now. That's mackinus in the face.

Speaker B: Yeah, that's uh, I feel your pain on that. We're like, I don't think, I don't think we push that button. But all right,

Speaker A: I think everyone's going through this stuff, right? But it's just like there's a lot of like after you get to the cool stuff. Okay, we need to roll this out to a few thousand people. That's, that is, um, that is uh, I have a lot of respect for companies that do that really well.

Speaker B: Yeah, um, couple more questions sort of as we turn final, final here. Um, one sort of, you know, you look at, you know, the 22, 2022 sort of version of you or you know, the team sort of coming out today, like what advice do you give them that you wish someone had given you?

Speaker A: I think it's the Paul Grammish type stuff that, that never really changes, which is stay unbelievably close to your customers, ship quickly shipped often, keep extremely high standards for the software that, that you're building, right? And then also, um, this is, I guess swerving over into, um, intel land. Um, but like, you gotta be paranoid. Um, and so you gotta, you gotta treat every day like you gotta just be absolutely paranoid. And so I don't say that in a bad way, but like, boy, there's just like you, you have to have such good market intel. Um, and part of that's talking that a big chunk of that's talking to customers. Right. But then also, like just being outside of your own four walls and seeing what else is going on just because there's so much capital flowing around, there's so many side bets being made, it's easy to be blindsided.

Speaker B: Yeah. Um, what have you seen from a sort of, you know, we talked a little bit about it, but like public sector versus your, like, large enterprise private sector. Have you seen. There's all these narratives that flow around on, like, the government's super slow or now the government's moving fast. Like, what's it, what's it been like for you? Because you guys sit at an interesting spot. You don't sit it like the, uh, hey, it's a, A, uh, drone that's like a cool video or, you know, a big factory. You're like, yeah. Hey, okay, so have you heard about data? Yeah, what the are you talking about, man?

Speaker A: I mean, what. So 12 months ago we were heavily tilted. Like we were pretty pre revenue and we're Silicon Valley terms where you're supposed to stay.

Speaker B: I've been told, yeah, yeah, you don't want more.

Speaker A: You know what revenue, they always want more. But we were pre revenue on the commercial side, right. We were about to launch on the commercial side and really start commercial being private sector. And then we had a bunch of sibbers and OTAs and stuff like that, right. That we were, um, establishing past performance and kind of understanding, uh, requirements and, and beginning to build with, um, army, socom, Air Force, you know, other organizations as well. What. And so we were heavily tilted in terms of ARR towards pub sec, but we had, we were still averaging, I think. I mean we have 70 million downloads of our open source right now. And I think we averaged about between 30 and 50,000 companies a week using it. And so we had like this large user base right. In private sector. But then we had this like, revenue on the, uh, on the uh, on the pub seg side. And over the past 12 months, we're now almost exactly 50, 50. And what has. I've been surprised in two ways. One, I've been surprised like this m. Like, been able to like, you know, pole vault the valley of death in some areas. Right. On. On the defense tech side of things with these customers and begin to scale with them in durable ways. But then also just like, how quickly we've grown on the private sector side. And for us, like something I'm really proud of, and I'm really proud of the engineering team here is we still have the same code base, same core code base that is running on Game Warden, um, as is running with our financial services clients or CPG clients or our, um, multi tenant SaaS hosted instance. Right. And so it's supposed to be.

Speaker B: That rocks.

Speaker A: We've tried to remain disciplined to the extent that we can on being a true dual use tech company. Same core code base, same core repos. Now It's a classified PowerPoint or it's other sorts of data flowing through it. Right. But it's the same core platform that we're deploying inside or outside of government.

Speaker B: Yeah, I love that. Um, all right, last question. Uh, because you've sort of already got this or seen a version of it, um, you're king for a day. You can change one thing. I'll always. I always use the, the example that Mamie gave because it can be about work or not work. Right. Mamie famously said, hey, I would have Chipotle have desserts so I could have a churro with my burrito bowl. And it is freaking. Just absolutely burned itself into my brain that I can't unshackle that. Like, I want a churro. So what up? Uh, what's the thing?

Speaker A: Um, okay, I mean, can I go with the, Can I riff off of that? Because I love it. I love, I think. Was it Al Davis who had the, uh, office in the back of an Applebee's? Um, I would love to have a desk set up in my neighborhood. Taqueria. I'm a, I'm a Mexican food guy.

Speaker B: I love that.

Speaker A: Yeah.

Speaker B: Hell yeah.

Speaker A: Um, but, um, but look, like in reality, you know, there's a saying that like, is kind of going around and that I've heard a lot of people say is like, you know, today are the good old days, right? Or right now is the good old days. And I'm having so much fun with our team and like, like what, what a time, right, to be doing this together and be doing this stuff that like, I think, you know, for me, like, having, you know, it's, it's hard to keep perspective on this, but it's also like, we all in this industry get to do some really cool stuff and work with some really fun people and some fun problems. And so, um, grateful for, grateful for

Speaker B: that opportunity, dude, that uh, that rips. And it's one of these where like, you know, you talk about sort of today in the good old days and then like you go back to the primer. Like I remember sitting out in the desert, freaking whatever, qaing or being like a sample customer for like primo primer demo like one or whatever. Yeah, totally. And I look now and you've got YC Defense thesis. Right. Like in the last month I've had the opportunity to like talk to kids at Berkeley, MIT and Brown. And if you told me 10 years ago that those four groups, YC, Berkeley, Brown and MIT were going to say, hey, we're going to like invest time, energy, people are interested in national security and defense, I would have been like, ready mind. And I do think there's just such an awesome kind of coalescence right now between really great technology, American innovation, ingenuity and sort of defense and national security and public sector that if we can harness it the right way, it's just going to have such an outsized impact and it's so much fun. Stressful, tired, all that.

Speaker A: But uh, I'd just say compared to maybe the worst of it like eight, ten years ago, with the depths of Maven and Google quitting Maven to where we're at today on the relationship between the Valley and defense tech, um, like

Speaker B: for some reason it does look like Google might be trying to quit again. But it's, it's still early. I think game one here, I think

Speaker A: folks that are passionate about the mission, like yourself, myself, like folks that are listening to this, like, uh, it's, it's a, it's a, it's a good. It's, it's, it's probably a high point right now, or at least close to it.

Speaker B: Yeah. Now this Rock's brother. Congratulations so much. All the success. You know I love you to death. I'm so excited just to see one, how great you guys are doing and two, just how happy you are as a great friend. And it's awesome. So thanks for being a, a friend, a partner, a customer, patriot, all of that. And uh, thanks for swinging by, brother.

Speaker A: Thanks for having me, Tyler. Appreciate it, brother.

Speaker B: Yeah, thanks everybody. Cheers.