From Services to Software: How Dual-Use AI Companies Actually Scale Inside Government | The Pair Program Ep94

Speaker A: Welcome to the PEAR program from Hatchpad, the podcast that gives you a front row seat to candid conversations with tech leaders from the startup world. I'm your host, Tim Winkler, the creator of Hatchpad. Join us each episode as we bring together two guests to dissect topics at the intersection of technology, startups and career growth. Welcome back to the PEAR program. I'm your host Tim Winkler, joined by my co host, Shawn Leahy. Sean, uh, you know, we just went through this daylight savings, uh, time this past weekend. Uh, I feel like it's always like that magical time of the year when the entire country just suddenly becomes a time policy expert. You know, everyone has this strong opinion on it. Like people will love it, people that hate it. Where do you kind of stand on the, on the subject?

Speaker B: Well, there's no excuse. It happens every year at the exact same time. I mean, that's the whole point of it, right? So, I mean, your cell phone, auto updates, maybe you have to change your wristwatch or if you have an old digital alarm clock. But, uh, I don't think this is just like people forgetting to pay their taxes in April. It's kind of the same thing every year. Um, but yeah, you're right, people always become these experts and then all of a sudden, uh, people start getting mad at farmers for some reason. Ah, but uh, yeah, it's a crazy time. But you know what, I feel like we're going to experience this again in 2027. That's my hot take. And my prediction for 2027 is that daylight savings will happen again.

Speaker A: It will happen again. It's like Groundhog Day. Yeah, I personally got a love hate with it. I love the extra hour of daylight. Um, but then trying to put a three year old to bed when her rooms completely all lit up, it is a bit challenging. So my position is we, if we keep daylight saving, you know, the caveat would be that government provides, you know, all the parents with some blackout shades. That'd be ideal.

Speaker B: All right, you know, you can write your congressman. That'll get, that'll get some votes.

Speaker A: Yeah, that'll get some votes, I'm sure. Um, cool. Well, uh, you know, speaking of, of time changes, it's, it's, ah, it feels at the moment like technology is going through its own kind of version of, of daylight savings. Now everything's speeding up, uh, especially around AI, new models, um, new startups, new capabilities popping up very regularly, almost on a weekly basis. Uh, but the real challenge isn't building the technology, it's, uh, integrating it into real operations. And that challenge gets even harder when we're talking about deploying this into government environments, uh, as we've talked about at length on the podcast. Uh, so today's episode is all about how technology actually scales in government, uh, and why, you know, some of the more successful companies that we've seen often evolve from services rolling into software platforms along the way. And uh, to unpack that, we've got two excellent guests joining us today. First off, we've got, uh, Darren Kamara, CEO of AI Squared. Uh, Darren and his team are tackling one of the biggest problems in enterprise AI today, which is the last mile problem. Uh, their platform focuses on embedding AI directly into operational workflows so organizations can move beyond pilots and actually scale adoption. Darren, thank you for joining us. Um, my pleasure. Nice. And alongside, uh, Darren is Dr. Be founder and CEO of Expression. Uh, Abeer Found Expression. I had to fact check this back in 97, uh, and has spent nearly, uh, three decades building software and data platforms for some of the most complex missions across Department of Defense and national security community.

Speaker C: Abir.

Speaker A: Uh, yeah, thanks for joining us as well.

Speaker C: Oh, no, thanks for having me. Really appreciate it.

Speaker A: Yeah. And what I love about our format, right, is that we've got two perspectives joining us on how innovation happens in this ecosystem. One from kind of the AI platform side, one from your mission engineering, you know, govcon side. Um, and so I'm looking forward to it. We're obviously getting more into, uh, the thick of it. But before we dive into the main discussion, we've got a little tradition here on the Pair program. Little warm up that we call Pair me up. So, uh, we go around the room, we spitball two things that just go well together. Um, Sean, why don't you, you lead us off?

Speaker B: Yeah, I'm coming in hot today with a, uh, technical pairing and one, uh, with a strong opinion behind it as well. My pairing today is AI and the terminal. Um, so I've been a terminal maximalist my entire life where I think that if you can't use the command line and you're deep into technology, you're lacking a critical skill there. That GUI is mostly just kind of gum everything up. And uh, since the release of Claude code, which everyone is using now to vibe code, you know, their own version of Salesforce or whatever, um, people have been like, relearning the terminal and falling back in love with it because it really is powerful and you can use all sorts of different tools and skills. So I think that uh, the future of technology is in, at some level, some degree, a return to the past and getting back, uh, deep into the terminal is the perfect pairing. The perfect pairing for AI.

Speaker A: Yeah man, I like that. Getting a little technical, get a little deep on it. I, I've personally have been able to like troubleshoot things that I never would have been able to do. So like going deep into command lines, uh, in my, in my computer, on just troubleshooting. Some, uh, some networking issues I've had. I uh, work out of the garage sometimes. We've got these little module antennas that connect from the house over to the garage. They went down and you know, AI just kind of helped me troubleshoot things that it would have taken me forever probably, or you know, getting a consultant out here to help out. But uh, I like where your head's at with that. It's, it's spot on for the theme of what we talk about in here as well.

Speaker B: Oh yeah, cool.

Speaker A: I'll jump in. Not, um, technical at all. Very, uh, different. Uh, my pairing this week is, um, bike rides and uh, proud parenting moments. Uh, so, you know, my daughter, you know, Alice, just turned three, uh, uh, just kind of starting to figure out how to ride her, her little bike. We got her one of these little balance bikes to start, so no pedals yet. Uh, uh, and since it's been nicer out, we've been going out around the block. Uh, she gets her bike, puts her a little stuffy in her basket. Uh, uh, she's got a little bell on, on the handlebar. She rings and it's one of these moments where I can see her each time, you know, getting progressively more and more comfortable and just getting a little faster and more confident. And I stand back and just kind of have this moment of, of a proud dad watching her learn. So she'll go like 10ft, you know, without wobbling or putting her feet down. And then I'm, I'm in the background kind of acting like she just won the Tour to France. You know, just real, real big accomplishments at this point. But uh, it's one of those, you know, small but big things, you know, as a parent, you know, watching their kid learn to, to ride a bike. So I'm going to go with your bike rides and proud parenting moments for uh.

Speaker B: This is, this is a very good snapshot of some of the dichotomies in the PA over here. Being a fantastic father and sharing moments of, of importance, uh, in raising his daughter and then me just yelling about terminals. So kind. Some different, different perspectives for sure. But I think it works out in the end.

Speaker A: You're spot on. Um, man, uh, let's pass around to our guest, Darren. Uh, you know, quick intro and, and your pairing.

Speaker D: Sure. Also, I'm Darren Kimoro. I'm the chief, uh, executive of AI Squared. Um, we, we bring, uh, AI infrastructure into operationalized environments, so we focus on things like data readiness, governance, user experience, etc. Um, clearly you both are very, very good at this pairing thing. I'm, I'm a novice at it, but the one that comes to my mind is a, uh, perfect weather and pollen. Uh, you know, we, we're now kind of getting into the longer days, as you mentioned at the, at the beginning here and uh, you know, wanting to experience it, spend the time outdoors yet, you know, at least here in the Bay Area where I live, this is also when pollen becomes like a really big thing. Cars become coated with pollen absolutely everywhere. So you, at least for me, uh, I've got, uh, allergies to the pollen. You really can't spend as much time outdoors as, as you would like, which may be a good thing. Now with vibe coding taking off by laws, me, or forces me for that matter, to spend more time indoors. And I would just comment on the concept of the daylight savings thing. I'm originally from Hawaii and Hawaii is one of the states that does not observe daylight savings. So for me it was a little tricky having to make uh, that adjustment. Uh, and unfortunately I do have a lot of those clocks that are manual and whatnot. So.

Speaker A: Yeah, the pollen is a problem. It's one of those things around here too, where today is a great example where we're staring at a 80 degree day almost and kind uh, of like this false hope that it's truly here. I've hold off intentionally on bringing out any like patio furniture or deck furniture just because I know it's coming. There's going to be a thick coat of pollen that's just gonna cover everything. So we usually wait off until a little bit later and then bring it out. But, um, yeah, you got to get that cleared and then you got, you got, you got a healthy Claritin dosage that you're, you're uh, you're sticking to, uh, at this time of year, you know.

Speaker D: Not yet. Not yet. But I have a Costco membership for that very reason to get my clarity ready. Yes.

Speaker C: Love it.

Speaker B: Costco levels of clarity.

Speaker A: Yeah. Necessary. Good stuff. Well, again, thanks for joining us, Darren, and ah, a beer. Quick intro in your pairing.

Speaker C: Sure. Um, I'm the CEO, uh, Of a company called Expression. We're private equity backed, based in the DC area. Um, provide uh, AI driven, uh, intelligence solutions. Uh, after the Department of Defense and other government organizations. My perfect pairing uh, would be autonomy and manual labor. Um, anthropic put out this graph on um, jobs that they feel won't be overtaken by uh, the recent wave of AI so in there being a plumber, an electrician, teaching piano, et cetera, et cetera. Um, so this is assuming that Elon Musk, uh, doesn't build uh, Zoptimus or whatever his humanoid robot is. It'll probably be 20 years from now before is that like FSD is like he's always like 20 years late on things. Um, but yeah, to me it's really interesting. Growing up it was always become an engineer, go into analytically based career sets because those will always be protected. Um, nothing's going to replace those skills in the human mind. And now we're at a point in time where that's not really true. Right. Those are the most easily replaceable, um, invocations. And then things that require um, actual manual um, manipulation are the ones that are most likely to survive whatever is going to happen over the next couple of years.

Speaker A: Yeah, it's, it's such a hot topic. It's something that we're obviously, you know, do a lot of talent on our side and so always kind of dialed into, you know, where, where are going to be those, where are those opportunities going to be that are AI proof or you know, a little less vulnerable? Uh, and it's yeah, folks building things with their hands, you know, folks that are, you know, getting into the nitty gritty of you know, a lot of mechanical types of uh, roles and electrical and stuff like that. But yeah, it's definitely, you know, something that everybody's kind dialed into and um, uh, it's a, it's a good pairing man. I think it's spot on. A lot of, lot of things that folks are, are yeah considered about but also understanding like how, how can it, you know, also um, you know, level them up, uh, but you know, without elimination, uh, entirely so. Well, good stuff. Yeah, good, good pairings all around. Um, let's jump into the, you know, the main discussion here. So today we're going to explore three main themes. Um, first, where innovation actually starts inside some of these different government missions. I uh, want to talk a little bit about how companies evolve from services into scalable software platforms and then what it really takes to move from pilots and prototypes into real operational Adoption, uh, solving this last mile problem in AI deployment. So let's kick things off with the mission side. Um, Abir, let's start with you. And the way I kind of wanted to start this and have it flow was give us a little bit of context, give our listeners a quick overview of what it is that you're building at expression and the types of missions that you're supporting. And then we can kind of follow on from there for some other questions.

Speaker C: Sure. So here expression, we, um, focus on the int. So if you're in the government market, you'd know what that means. So there are the varied, uh, intelligence subdomains. Um, for about the last 20 years, we've really supported SIGINT, which is signal intelligence. So anything and everything to do with the spectrum. We, uh, built one of the first iterations of a battle management system for the DoD, uh, to handle, um, spectrum management globally for the US military, um, and spectrum operations. And we took a lot of that base and we built out our. We have an agent AI platform. Everyone has an agentic AI platform these days. Um, ours is specifically focused on the intelligence market. So, um, we take in signal intelligence data, make sense of it, help operators, uh, with cognitive, uh, fatigue, looking at all this information coming in worldwide, making sense of what they're seeing and then giving them options. We do that in the human intelligence side. Um, if you ever have a security clearance, uh, our platform is actually utilized to perform human intelligence across the whole DOD industrial base. Uh, so actively our agents are looking at your financial information, your travel records, um, who you're connecting with communication, who's in your circle, who's in their circle, how far are you removed from bad actors, et cetera, to be able to make continuous evaluation on your worthiness, uh, for your security clearance, um, we do that across, um, supply chain risk management, uh, for the dod, um, really in all these places where, um, there's a large volume of information needs to be analyzed, uh, um, uh, you need to make sense of it and then understand how it relates with your mission set. Those are the places that we tend to excel.

Speaker A: Nice. So just from a, you know, kind of like a bird's eye view here, um, your perspective specifically when it comes to, you know, AI innovation, let's just say inside government environments, where do you kind of see, you know, these real operational problems typically start to surface?

Speaker C: Well, I mean, a lot of the issues are the, you know, the foundational models aren't quite ready yet. Right. So if you're talking about warfighter functions, you know, we're still behind fully, um, autonomous, um, decision making, uh, decision support, um, hallucinations. You have um, biases in training. Um, most of these models are trained off of a lot of open source and maybe not so open source data. Um, and so uh, the problems we see are that um, those models are only as good as the data they're trained on and only as good as the scale they can be utilized at, um, their emergent properties in AI. I also am a professor at Cornell University, so actually I teach a lot of this to my students. Uh, emergent properties that we don't understand, but that all has to do with those scale and good information coming in. Um, and so um, on the government side that's a huge issue when you're using it in uh, war fighting functions. Um, you have to be able to guarantee that those pieces exist. So there's a lot of AI human teaming still speeding up things. Um, but the human in the loop is going to be there for a bit until we solve those problems.

Speaker A: Yeah, and I want to pull on those threads, um, a little bit later on here in the conversation I want to quickly pivot over to Darren and do the same with you. Just for AI squared, give the listeners a little bit more of a detailed overview of the platform and the problem that you're solving and then I've got a follow up for you.

Speaker D: Sure, absolutely. So our, our company was started by a group of NSA technologists. So they had initially developed the technology at the nsa. Um, what they were trying to do is create a secure environment to be able to deploy models, uh, at that time, machine learning models predominantly, uh, process highly sensitive information, generate summaries and then move the uh, reports around in a secure way. And that for all intents and purposes became AI squared. So what we do today is we take our data platform, we connect it to all the different sources of data an organization might have. And it could be any varied sources of data from things like uh, databricks or Snowflake, uh, into S3 buckets or uh, even into SaaS applications. We ingest that, bring it back into our platform. Uh, in our platform you can run uh, traditional foundational models but also bring your own models. In some cases our customers have created their own fine, uh, tuned models or even small language models. You create workflows within our platform and you connect those into things like governance modules which review uh, policies. Maybe it's nist, um, maybe it's some fars. It uh, could be HIPAA or whatever the uh, applicable policy, uh, governance that you're trying to observe is uh, and that allows you to ensure that the results coming from those models are compliant. And then we also add on some additional capabilities such as things like masking so we can hide uh, pii, for example personally identifiable information. Um, we have feedback loops where we can actually observe how people are interacting with the results and incorporate that into uh, tuning the model for the next version, improving the model. So the models are constantly learning and getting better. Um, so that's what we do. And then what we do at the very last mile is we take the insights. And sometimes they're dashboards, widgets, charts, graphs, uh, sometimes they're just like chat queries and results. And we can uh, create UIS dashboards for our end users or more importantly we can put them into applications that they already have. So in the case of uh, the Department of War work that we do, a lot of this is augmenting dashboards. They're already using, perhaps they're using some uh, uh, existing dashboard, uh, tableau, a qlik dashboard and they want to get a little bit more information. What we can do is we can actually augment the window that they're in. We slide all the information from that particular dashboard to the side and on the other side we actually have the AI generated information. And that's going to be context aware so it's going to understand the information on that screen, but it's going to pull data from other sources. You uh, know a bunch of randomized sources, run uh, it through the models and then surface the information so that you can have a little bit more context, a little bit more information about whatever it is you're taking a look at. So it gives that user a lot more uh, power, a lot more information and then the ability to interact with the data. You can ask the data, uh, questions, you can chat with your data, dive in a little bit deeper than what you see in that particular widget by itself.

Speaker A: Quick shout out to our sponsor, Defense Unicorns. This one's for the problem solvers out there. They're hosting Warhacker, a uh, first of its kind hackathon built for the defense community. No buzzwords, no slide decks, just hands on keyboards solving real mission problems with real code. You'll be side by side with developers, engineers and innovators from across government, industry, nonprofits and academia, all hacking for the warfighter. It's happening June 16th through the 19th in San Diego. Got a real world problem to solve or want to join a team that does? Learn more at defense unicorns.com hatchit yeah, it's very cool. I want to kind of pull a bit on that last mile problem that you, you, you touch on there with you know, this gap between building the models and then actually deploying them into these real workflows. What are some of the biggest obstacles that you kind of see when organizations are trying to operationalize some of that AI?

Speaker D: Yeah, the biggest one that we see is um, when you're building one workflow for one end use application a um, lot of times what we find is organizations will try to stitch it together themselves. M In many cases there's no roadmap for them to follow so they're doing it by trial and error. They're trying a bunch of different tools. Um, the problem with that, so over time uh, you'll get it to work, you'll figure out the right tools, you'll be able to make things connect, um, you'll be able to write the code between each of the different components and now you have a viable workflow. The challenge with that is that it's not a durable workflow. What we have seen over time is as uh, SDKs Change or perhaps APIs change or data in the APIs change, the uh, the workflow itself becomes unreliable and in some cases they could even break. And the challenge there is then you have to go back and figure out how you constructed that original uh, workflow. And in some cases the person who built them, they may not be around anymore, they, they may have forgotten. There's no documentation because they were building it on the fly. And, and as a result um, it's hard to you know, replicate it, it's hard to fix it, it's hard to you know, basically it erodes trust at the end of the day because the end user is no longer confident in that particular uh workflow. So what we're trying to do is we're trying to institutionalize it with uh, SLAs, um, with digital twins to ensure that we can test environments before things drop. So that way it's not going to actually impact the end user in the real world production environment, um, and make sure that they're always up. So uh, we're instrumenting that entire workflow from end to end with agents to be able to provide alerting and monitoring. Um, so if anything does drop we can get on it right away and bring it back up as quickly as we can.

Speaker B: So um, Darren Abeer, you'll find out pretty quickly that I function as Tim's um, live GPT for some of These episodes right, where I'm constantly ingesting context. Um, and a theme that I've detected already is um, cognitive load and human in the loop. You actually both use those terms. Um, and uh, I think that we're seeing with this sort of AI uh revolution that the speed and velocity of development of software applications, of models, both new large language models and more traditional machine learning uh, models is incredibly quick and you can get access to a ton of data and you can analyze that data now faster than you've ever been able to do it before. But is it actually leading to faster or better decision making? Because ultimately, especially in the, in the realms that you're operating in, highly classified defense and intelligence related, there's always going to be a human in the loop. I mean Abir, you said that in some of your initial remarks is that there's, there's going to be a human in the loop for uh, at least the foreseeable future. Um, in your experience has there been any sort of like cognitive overload that you've seen in deploying these, these workflows and other agentic solutions where um, the human in the loop is actually now also the bottleneck? And uh, I'll start by directing that back to Abeer but then kind of the same question for Darren after Abeer's comments.

Speaker C: Yeah, we have seen right. So some of the um, kind of structural problems um in the federal government tends to be the training of the workforce or the composition of the workforce. Um, so we have a cyber use case uh that um, are in the process of deploying DHS wide. Um so there's something called a security Operations center or soc. Um, and so our gendic platforms taking in data from these SIEM providers, Splunk, uh, elastic couch, CrowdStrike, et cetera. Uh, and then um, um, looking at the events of interest and building patterns of life, you know from those events like ah, at a very pared down level to push this to soc analysts. You know a lot of the SOCs are manned by junior analysts, um and so they're unfamiliar with understanding what patterns might be a precursor to a larger cyber attack or what have you. Um, and so the AI is able to understand those patterns of life very quickly. Right. And determines these five events actually equate to a certain pattern of life and then that's being repeated hundreds of times. So danger, danger, um, and those junior SOC analysts because of cognitive overload, this is one of terabytes of data coming in daily, uh, even though the AI is able to alert them to It a human has to then action some defensive measurements or what have you and they're still being overloaded. Right. And our AI is prioritizing what's the most critical risk to infrastructure, especially on the cyber physical side. And still the personnel, they don't have the personnel to be able to action that off the AI. So you still need to have humans that are well trained, um, and enough of them get in the flow of this information in your domain.

Speaker B: Yeah, that's uh, I also like that you specifically kind of pointed out some of the more junior folks. Right. If you're, if you're more mid level or senior, you have developed some muscle memory and some tacit knowledge. But if you're just starting out, you're still getting your head wrapped around things and you have these AI agents that are just uh, kind of outpacing you. Uh, Darren, I wonder if you want to jump in here with similar perspectives on what you've seen.

Speaker D: Yeah, so I think this is all use case, um, specific, uh, in my opinion anyway because um, I can see situations where you will have more cognitive overload, like the SOC scenario. I see a lot more, at least in the work that we do where it's the other way around. We're actually relieving, we're providing cognitive relief. And I'll give you an example of what I mean there. So in our work we do a lot of work in governance. And uh, a lot of times what happens is uh, in order for you to be truly applying the appropriate governance, you have to be uh, using the state of the art, the latest and greatest policy, uh, documents. So it could be starting from a NIST standard, but then there could be a series of memos that have come after that which uh, provide additional guidance and restrictions on uh, directions or rules or applications, whatever the case may be. And where we help is with the analyst, the analyst who's responsible to review policies that are coming out from a particular department and they need to be compliant with all the other departments that relate to that particular subject matter. It's hard for one person to know that without using AI. AI allows them to very quickly review or write uh, something, a document for example. And the AI has within its corpus of knowledge all of the uh, related governance, uh, materials that it should be scanning against. So as you're writing or even after you've completed a report and you run a scanner against it, it can instantaneously highlight areas where, yes, this is green, Great, very compliant. This is yellow. Perhaps something you should look into and footnote it click on the footnote, go to the sidebar, see exactly why it could be questionable or perhaps even read where yes, this is just wrong. And AI can actually help suggest us uh, alternatives rewording uh, for example to make sure it's not as ambiguous uh, for that particular user. So this allows that analyst to be able to process a lot more information a lot more quickly and a lot more accurately.

Speaker B: So this is really interesting. I'm glad that we kind of did it in that order because I think that my initial takeaway is um, for you know, agencies, organizations, federal, but also corporations commercially. When you're trying to find ways to deploy AI, obviously there's a lot of different landmines out there to avoid. But so much of it is understanding a cognitive loading or a cognitive de loading. Right. And so, and what Abeer pointed out, that could be some cognitive loading that you're going to have to have your workforce be ready for, be trained up on what have you. Whereas Darren, from your experiences there, there's also ways in other use cases where you can cognitively deload. Um, so for tech leaders out there it's you know, easy problem. Right. Just completely understand top to bottom your, your problem space. Um, but just that's interesting stuff that you don't really see in a lot of the uh, kind of normal content commentary on AI deployment. But um, Tim, I think I'm going to push it back to you and you can keep us oriented here.

Speaker A: Yeah, I wanted to um, kind of jump into this theme that you know, we've been seeing, you know, a bit more and more of in ecosystem which is kind of moving from services into scalable software platforms and not really moving entirely from services, still keeping services, but tacking on and adding on um, you know, a software platform to your, you know, all in one solution if you will. Um, Abir maybe uh, I'll kick things off with you, you know, with, with the platform or the, the tool that you all have been building. Um, you know, what were the patterns across the programs that you were supporting that made you realize that it was time to build a product rather than continuing purely as services. Was there certain patterns that you're kind of picking up on that said, this would be a much better outlet for us?

Speaker C: Yeah, certainly. I mean a lot of the government problems are um, within uh, one or two classes of general problems. Right. Um, you look at, um, especially on the Department of War side, you have a lot of legacy systems that have been stowed pipe, but they support a domain. Um, you need to Bring them into a single context. The Accenture, the Deloitte way would be to bring into a data lake what have you, but you still don't have semantic understanding. So a lot of what we've done in the domains we've worked in is take that disparate data, um, but they all support a single domain, um, and then uh, we give a semantic understanding. And once you do that actually um, deploying agents becomes quite easy. We see uh, an old consulting model bringing the data lake in there, you're doing rpa. Uh, but really in today's world, given the, the power of the foundational AI models, really what you're looking to do is bringing that data into some context where you have semantic understanding and then uh, giving it access, giving those foundational models access, whether through uh, agents or directly, um, to be able to expose that knowledge base back to the operators. Uh, and then once you do that piece, then you look at all these different problems that exist in a domain and they're very self similar. On the intel side it's you know, uh, am I, you know, targeting, am I trying to figure out a pattern of life? Uh, you know, um, you know, and you know, you're repeating that whether you're in cyber, whether you're in, you know, signals intelligence, whether you're in human intelligence. Um, and so you know, though, you know, you could do kind of anything and everything, uh, in the agentic lane, you know, we've stuck with where we as a company have done well traditionally and then just have doubled down on the AI piece.

Speaker A: Yeah, it sounds like a, certainly repeatable, just different domains, uh, and program areas. And then I guess Darren, on your side, maybe approaching this a little bit from a different angle in terms of embedding AI insights directly into these existing applications and these workflows. Why does kind of like where AI shows up matter so much for adoption?

Speaker D: Yeah, and I'd like to maybe take a stab at the previous question as well because I think um, our journey was interesting and it may not be that dissimilar to other um, listeners to uh, this podcast. So we started off basically focusing on building technology through SBIRS and CTRS and we did over 10 different government grants and really the goal was to create a bunch of technologies that could be scalable and usable in an ideal dual use kind of environment. So based on that you would think that we would have ended up with a really great technological, ah, product that could be highly replicable in its use and scale from that going forward. What we realized though, in the reality of it is that AI, uh, is very, very immature in everything that it is. I mean, we've seen foundation models continue to mature real time before our eyes. Um, the environments that are, you know, where we're trying to deploy these, um, systems are not AI friendly, they're not, they're not easy to deploy AI in IT yet. And again, as I mentioned previously, there's not really like a playbook yet that allows you to just do this, then that and equal success. Um, I think those things will come through maturity, but they're not there yet. So what we found is that as a technology company, we had to become a little bit more of a services company where as we're bringing AI into these very different, uh, IT organizations, they're constructed in completely different ways. They have a bunch of different tools that don't talk, they have systems that don't recognize each other. Um, even authentication is difficult in these environments. And we actually have to go in and build the connectors. We become a little bit more services. Like now we try to always do that in a software way where, um, you know, we're not trying to sell, you know, professional consulting hours. We're definitely not trying to become like a Deloitte or a Booze or anything like that. But, but that's required in order for you to have the success, successful deployment of AI. And what we're finding is that even within those organizations, and I think this is maybe the answer to the latter question, uh, you still have to fine tune these models. The foundation models are great, but they're very generic. Um, they won't provide information that will specifically address how the organization is functioning today or the types of decisions that they're making. And in fine tuning those models, we can actually help kind of fit the model to the way they work as opposed to making them work to the way that the model thinks. And that's a lot more adaptable for them. So what we found in doing that is that you get a lot more buy in a lot faster because they're not changing their workflows. The models are changing for them. So we're bringing AI into their existing environment.

Speaker B: I think, Darren, that's, that's hit on something that's really, um, important because it's, there's some historical antecedents to it. Um, in years past, if you wanted to bring a new technology to the federal world, um, you couldn't just show up with it and say, hey, isn't this a cool new technology? You had to understand their mission, you understand their pain points. Um, and if you didn't do that, you were never really going to get traction. And what you've just outlined is kind of the AI version of that, which is you can't just show up with a couple of, like, clawed skills and, uh, you know, just throw it at whatever data that exists in an agency. You really have to understand how their organization works, how the organization thinks, and then, of course, the mission sets that they're executing, um, and to then craft these AI systems that are actually going to be useful. So I like that you were honest there and said that AI does a lot of things really well, but also it's still pretty immature and there are some things that it doesn't do well, um, that an experienced practitioner can call out immediately. Um, but you have to build these human machine interactions, um, effectively. If you don't, it's not going to, uh, it's not going to go well.

Speaker A: Oh, go ahead, Darren.

Speaker D: I was just, I was just going to agree. I mean, that's, that's a great way to summarize it. I, I think that's where we are in the maturity. I think as we, you know, as society become much more comfortable with it and we begin to, you know, adapt it more into our, the way we go about doing our business, you know, things can change. But that is definitely where I see it today. That's definitely where we see the Department of War today.

Speaker A: Yeah, I mean, I, I can just speak, you know, firsthand from, you know, supporting some of these, uh, efforts from a talent perspective. And you're looking at these government systems that are, you know, decades old, you know, still, you know, mission critical. But the idea of, you know, you know, rebuilding them entirely, it just seems, you know, unrealistic versus augmenting them. It seems like a much more practical approach, um, obviously moving faster as well. If you can enhance that existing infrastructure, uh, instead of replacing it. Um, I, uh, wanted to kind of ask though, on this kind of theme that we're kind of seeing, you've seen companies like Palantir also build hybrid models, um, part platform, part services, kind of deeply embedded with the customer. Is that kind of the future of govtech companies, do we start to feel like that's going to be what the majority of these types of companies look like, is this hybrid model moving forward?

Speaker C: In part. Right. I mean, the Palantir model is interesting because Foundry was, um, largely a collection of different tools that Ford deployed engineers built. Uh, and you needed engineers to configure them for use. Right. I, um, Mean even to today, uh, you need those engineers to. Nothing that Palantir does is out of the box. So you need those forward deployed engineers to configure things, to build data pipelines, to build ontology and then you need them to operate. Right? So it is really a services company masked as a SaaS company. Um, I think with the foundational AI, uh, model providers, um, uh, it would just be like Gen 4, Gen 5 of programming languages. Right. So back in the day I'm old enough where I loved fractals as a kid and I actually built ah, and I didn't have a computer with a math, ah, coprocessor. So this is really back in the day. Uh, so to generate a fractal, um, I built my own assembly library to do fixed point math, uh, that you could use your registers on your CPU to quickly, quickly run through. So I had to understand how the registers worked and the underlying properties uh, of my cpu. But now with quad code, what have you, you can largely make natural language prompts and then it uh, interpolates that into code that then becomes an application. Eventually it will just be the natural language that goes directly uh, to uh, to your program, right? So it'll create uh, behind the scenes machine code that someone uses. And then when that happens, um, I think the model is going to be very different because you can have a different application every day of the week. You can make those modifications as you like. You're not stuck with your Salesforce or your foundry or what have you. As long as the data exists somewhere you can build, everyone can. The four of us could be in the same domain but have a customized application to our liking. And the O and M and everything on that is low because the model's generating all that for you. And to me that's actually the future. You won't see what you see today. Uh, this is the hybrid state, right? And the future future state is that you prompt the model for whatever you need. It generates a UI that's most suitable to you and that's how you interact with data and operations in the future.

Speaker B: So uh, Abir, you're speaking my language here. I started off the episode talking about CLIs and terminals, and now you're talking about writing your own um, basically compiler it sounded like uh, back in the day. So this is fantastic. Um, and I completely agree with and love your idea of characterizing these large language models as generation four, um, programming languages. Right. You know, you can think of an LLM as a compiler that doesn't have syntax Errors. Right. You just, you throw in natural language or even incomplete natural language. I always notice that my spelling gets worse when I'm, when I'm talking to Claude. Um, but I'm gonna, I'm gonna not push back here, but I'm gonna play Devil's Advigar. I'm gonna red team this a little bit, which is. Yeah, I mean I can, I can now use natural language to develop dashboards, interesting data analysis and even like full scale applications, uh, to my heart's content. But if I'm mischaracterizing the problem or if I have like a flawed ontological understanding of what I'm looking at, if I don't really have a good kind of, um, source of truth, is there a danger in the feeling that I'm like gathering insights and making better decisions versus the reality of I'm kind of um, going down a rabbit hole of my own making?

Speaker C: So critical thinking always has a premium in life, right? I think for.

Speaker B: Oh God, I was hoping you wouldn't say that. That's not good news for me.

Speaker C: I'm sorry.

Speaker B: Continue, Peter.

Speaker A: You're sure?

Speaker C: Uh, so there's a time there, there are all these great memes about programmers. Um, back in the day you'd be an applied math or a physicist and go into programming. Computer science became its own domain, um, and properly trained computer scientists have a lot of critical thinking skills, you know, very defined understanding of logic, you know, predicate logic, Boolean algebra, so on and so forth. Uh, and then we came to an era where programming was relatively easy and you know, you could uh, you know, go to a code camp or what have you and you know, and understand how to program. I think we're sort of coming full circle back where, you know, when this fourth generation, uh, programming language via the LLM truly is, um, viable than critical thinking skills. How you deconstruct problems, how you uh, understand decomposition of problems is effectively prompt engineering. The better you can decompose a problem, the better you can generate prompts, the better the LLM can build for you. Eventually there will be guardrails, so you can be very lazy. I need a dashboard to tell me how the weather is going to be for the next whatever. Right? Um, but that's still far off in the future. The compute that requires is fairly extreme, at least based on how models work today. So we're going to have a golden era of critical thinking, uh, and decomposing problems, defining problems, uh, and that will be sort of the next era of the software engineer. Right. Those who can do that. Um, will be able to leverage this tool to its full benefit. And you probably have organizations instead of having forward deployed engineers, you'll have these, um, thought leaders, uh, in domains that can take a domain understanding along with, um, critical thinking, bring that together and then they provide the guardrails to your rank and file. Um, and that is, um, I guess, sort of like the hybrid model, but even more specialized.

Speaker B: Yeah.

Speaker A: So as a professor, is this something that you're bringing into, uh, your coursework is like focusing in a lot on critical thinking?

Speaker C: Well, so I was probably the last class of people to get their PhD before everyone used AI and I had to teach. During your PhD, you're forced to teach. You're an indentured servant to the university that uh, you obtain your doctorate from. Um, and so as I started teaching post PhD, I do it on Mondays, two graduate courses at Cornell. Um, I started noticing my students using AI to answer problems and it's spy versus Spy. Right. So how do I generate problems that require critical thinking that you can't prompt completely? You need some context and understanding. Um, or the AI is going to give you gobbledygook. Um, and it is something I teach both directly. Um, like hey, I'm giving you problems that will. Won't produce the appropriate answer if you just take the PDF and jam it into Claude or into um, ChatGPT. Uh, and it's sort of, um, in a backhanded way teaching these guys to critically think. Um, they're from a generation now where they've gotten to use AI to solve most of their homework. They're not learning foundational skills. And so when they get to the graduate level, um, you know, they're really at a loss. Right. They had Covid, now they had AI. So they had just this huge gap. Uh, and it's a problem on how to teach them to think critically, where, you know, they didn't interact socially for four years, you know, with others, and now they have the easy button. We used to have Cliff Notes, right? Cliff Notes were sort of great, but you still had to read them. You know, so Instead of a 300 page document, you had a 50 page document, but you still had to read those 50 pages and you still had to write your own essay. Um, and they don't have any of that. Um, I think it's a problem for educators, uh, by and far because, um, kids are very industrious and they're able to use the AI to probably you could do a year's worth of schooling in a week's. Worth of time and get straight A's. So

Speaker A: it's such a good point. Yeah, it's such an important, important point you bring up. It's actually something I think would be a great episode that we craft, uh, get a couple of professors, maybe, you know, get you on a repeat episode here for a future, uh, you know, episode on critical thinking of beer. Because I agree with you wholeheartedly. I think it's just, I love the easy button, uh, you know, analogy, because it's so true. Um, Darren, I want to kind of, you know, kind of wrap a little bit of this main conversation and lean into, you know, scaling because this is something where we get a lot of early stage founders that join, um, uh, as listeners here and building something, you know, especially if we're talking about dual use and they're trying to, you know, maybe break something into, uh, you know, these more regulated industries, if it be defense, national security or what have you, um, you're going from that, you know, that prototype or that pilot into, you know, real operational scale. Kind of talked a little bit about it, but maybe just reiterate a little bit on, you know, when organizations are trying to, you know, move from that, you know, prototype phase, you know, what typically kind of breaks, you know, or where do you see, like these biggest bottlenecks? You know, if it's integration, you know, uh, ready, organizational readiness, you know, can you just expand a little bit more on that?

Speaker D: Yeah, you're happy to do that. And just a comment on the previous topic. So I'm a PhD student, so. But much later, I've got another year.

Speaker C: Where are you, dud doing your PhD?

Speaker D: I'm at Claremont, uh, Claremont Graduate University on the West Coast. The la. Yeah, yeah, yeah. So it's funny to hear the other side of the argument because as a student, you know, I get to, uh, be the beneficiary of some of the advanced technologies and, you know, don't have to read the 50 page cliff notes anymore. So it's really interesting topic. Uh, you just teed us up for

Speaker A: the next episode then. We already got this. We'll bring both of you guys back. We can go again.

Speaker D: So I will say with regards to moving into the production, um, you know, crossing the cavern, if you will, uh, you know, and trying to go from pilot to production. It is, um, very, very tricky because the different environments have different requirements and different organizations also have different expectations. What I would probably say is some of the areas we see things typically fall apart, um, is a couple things, I guess, that Come to my mind. Uh, and these are more maybe just observations, uh, and perhaps even some tips. But um, when you move into the production the pilot environment's pretty easy. Pretty much anything goes. You're in a lab, you're not connected to any production workloads. If something goes wrong you're not going to really harm anybody. It's isolated and protected. The production environment is a completely different animal. If a uh, result, if a query comes out and it's uh, providing instructions which you know, in the particular in the case of the Department of War, um, it could lead to loss of life, right? It could lead to loss of multi millions of dollars. So having um, you know quite a bit of cycles on, on the technology, having the um, uh, the users, uh, similar to I guess when I think about in analogies like Waymo learning the streets and having to um, be able to produce a certain amount of hours before it can actually drive on those streets autonomously. AI is very much like that. You have to be able to produce you know, replicated consistent results uh, before you will have a human accept that and push it into production. So I guess the first thing that comes to my mind through my experience here is it's going to take a long time um, it takes a long time to move into the uh, you know, the production environments, the high il environments for the department uh of War or into production environments for the, for the enterprise. Um one thing that always falls apart is cybersecurity. I think that um, oftentimes when you're building in, in pilot environments you tend to want to go very very fast and you tend to forget uh, about some of the requirements that uh, from a cyber perspective that you must have in, in the production environment and, and gaining those um, doing the appropriate scans, having the right technologies which are themselves ato, um, very very different because again they're typically expensive. Um, they're, they have their own requirements for implementation. I know when we moved into the production environment, uh, you know I got hit by a bill that I was not expecting because we had to build by a number of different technologies that provided um, the minimum requirement for meeting nist. But we didn't care about any of that in the pilot environment. Um, things like sbom, producing software, bill of materials that could be reviewed line by line. Um, that's not something you typically do in pilot environments. But you have to do that and someone has to review that and review your open source code. Uh, and someone at the end of the day is going to be responsible for signing off on Those things, um, before it goes into production. So I think you have to be prepared for that. Um, there are companies that can come in, companies that come in and help you for example pen test your software. You could have cyber security reviews of your software. There are tools that can help you. But if you have never done it before, uh, you know you, it, you have to probably plan for the worst because uh, I can tell you uh, again from my own personal experience, we had thought it was going to take a particular you know, period of weeks to get through it and it ended up being much, much longer than that, over six months to actually get through it. Of hard work and focus every single day to get us finally uh, into production.

Speaker A: Yeah, super, super helpful feedback there. Um, I want to you know, make sure I get a beers kind of you know, 2 cents on this as well and we'll then kind of wrap the main discussion and close with our final segment.

Speaker C: Yeah, no, I mean the, I mean to Darren's point, right, like as we deploy into you know, different DOW environments, uh, the you know, unlike you know, commercial deployments, when you support uh, the US military, you know there are these different classification levels, all these cybersecurity related uh, requirements and a lot of them are legacy. They could use AI themselves. Uh, uh, Darren mentioned to uh, the NIST like AI 853 standard. So those are all largely manual controls via the risk management framework. Yet you get uh, real time telemetry on all your systems as well. Right. So there's the opportunity for DOW or for the government to have these digital twins uh, that could accurately describe the state of your cybersecurity posture. But all of these are again stove piped. They were built independently and not with this common uh, thread so to speak. Uh, and because of that there are all these, there are costs that are borne by companies like ours and Darren's um, because of inefficiency and lack of cohesive policy understanding and stovepipe systems rather than um, what would be best practices. It's something that we probably learned from our near peers. If you go to your China, um, you do your Chinas of the world because they control the industrial base to uh, your military deployment, uh, they're completely aligned. There's no daylight between what you're doing from industry out to what you're providing um, for military or other operations. Uh, and that um, capitalism is great but that type of cohesion, uh, and having an end to end, um, like having uh, people end to end work lockstep, there's a value to that, that we currently can't replicate. But I imagine, uh, you know, in the next, you know, you know, next decade, we're going to have to get to, because that those inefficiencies are where, you know, our, um, our adversaries. Right. Can take advantage of us.

Speaker A: M. Yeah, well said. Um, I had some other questions I'm gonna, I'm gonna pause on it because I, I know that we're running up, uh, against the, the hour mark and we want to have some fun with this last segment. So going to put a bow on the, on the main conversation. I think we got a lot accomplished, uh, with, with the time that we had. And we'll, we'll segue into the five second scramble. So this is a fun one. This is, uh, rapid fire Q and A, kind of say the first thing that comes to mind. Try to keep it within five seconds. Be some business, some personal. Not, um, too personal. Sean, why don't you, um, lead us off with Darren and then I'll close with a beer.

Speaker B: Happy to. All right, Darren, five second scramble. Easy. We'll start out with the, uh, first question. What is something about the culture at AI Squared that you want everyone to know about?

Speaker D: Fun? Do you want me to elaborate or is that good?

Speaker B: Great answer, great answer. Less than a second, three letters. Outstanding. Um, because AI Squared is a fun culture, uh, what are some of the, uh, roles that you are hiring for currently or plan to in the immediate future? This is so that some of our listeners who might be interested and excited about working with, you know, what you're looking for.

Speaker D: All of them. Sales, technical data, science. You know, we're hiring pretty much across the board. Yeah.

Speaker B: Okay. Outstanding. Um, all right, what's the best decision that you've made under serious time pressure? Wow.

Speaker D: Serious time pressure. I don't think I've ever made a great decision under serious time pressure, honestly.

Speaker B: Valid answer, valid answer. Um, what's a skill that you thought was useless when you learned it, but it turned out to be very, ah, valuable later on in your career?

Speaker D: I think presentation skill. Uh, I never really thought having to present, you know, in school you have to present and all the boring. But what I found later in my career is that that's probably the most important thing, you know, being able to take a message, make it clear, make it efficient for a listener.

Speaker B: Awesome. Um, yeah, and very common, especially amongst more technical folks, is thinking that that sort of skill, the softer skills, presentation skills or afterthoughts. I can see Tim is gritting at that. Um, what's a failure that taught you something that you still use today.

Speaker D: Resilience. I think, uh, you know, oftentimes we, we have the best crafted plans and they never work out. But I think what you learn from that is uh, it's how you get back up. Right. And what you do with that learnings, you know, repeat it, that idea, learn from it, improve on it. Uh, you know, that's, that's what I try to do every time.

Speaker B: Awesome. All right, you made it through the hard part. The next couple of questions should be a little bit easier. Uh, what was your first job ever?

Speaker A: Electrician.

Speaker D: Uh, working for my dad.

Speaker B: Did he make you test to see if the current was still going through the sockets?

Speaker D: Yeah, like this ah, guy climbing in the roofs, uh, where it's tight. And I was a kid so I was the only one who could get up there and wire those little light fixtures at the corner of the house. Yeah,

Speaker B: that's a good role and task match right there. Um, if you weren't in AI, you know, assisting the defense world, what industry do you think you would want to disrupt?

Speaker D: Uh, energy. I think it's a passion of mine. I think we have tremendous energy problems, clean energy in particular. And I'd love to continue to help figure out a way to solve that problem.

Speaker B: Great.

Speaker A: Very top of mind.

Speaker B: What's your uh, go to stress food or stress drink? When you're deep into a project, when you've got a deadline coming up, it's midnight, what do you go to the fridge to get to keep you, to uh, keep you running.

Speaker D: Oh, chips. I'm a big chips guy. Any bag of chips I can find, I'm eating it at all of it.

Speaker B: Do you have a flavor? Uh, do you have a favorite chip flavor?

Speaker A: Yeah, throw your favorite out.

Speaker D: Personal guys. I love burritos. You know, the artificial everything types of chips. Yes. Spicy hot.

Speaker B: Yeah, Very, very popular. Um, now your best podcast or newsletter recommendation that has nothing to do with tech because we know that that would probably be the pair program. So we want to give everybody else a chance. Non tech related podcast newsletter that you think is, ah, that people should check out.

Speaker D: Well, I listen to the Journal every day. I think that they do a nice job of summarizing one big current event story. And you know, it's very short. It's like 17, so that's my goal too.

Speaker B: Excellent. And then final question. It's our standard closing, uh, question for the five second scramble. Uh, what is a corporate philanthropy or charity Charitable organization that's near and dear to your Heart that you want our listeners to know about?

Speaker D: Uh, you know, I'm a eagle scout, so I, I support the Boy Scouts of America.

Speaker B: Fantastic. Uh, Darren, you've crushed it. Five for five on the five seconds grab. Actually ten for five on the five second scramble.

Speaker D: Thanks.

Speaker A: Um, ten for ten actually. All right, let's, let's jump in a beer. You ready?

Speaker C: I'm ready. Let's go.

Speaker A: All right, if Expression were a restaurant, what would be its signature dish?

Speaker C: Maybe like a 72 hour pho that was like lovingly made by a series of grandmas in the back at it, you could see through like a glass pane window.

Speaker A: Oh, that's, that's fantastic. It's a good visual too. Soothing. And a good hangover. Exactly.

Speaker C: Great. Good grace for the winter.

Speaker A: That's right. Um, what, uh, are some, some active openings that Expression is hiring for over the next three months?

Speaker C: Infrastructure engineers, uh, AI engineers, uh, and then BD execs. So you know all parts of the sausage, right, like uh, making the sausage, selling the sausage, and then making, uh, sure that you quality control the sausage.

Speaker A: Nice. What's something that would surprise candidates about the culture at ah, Expression?

Speaker C: Um, obviously we're nerds, but we're competitive nerds. So we have a lot of uh, trivia, all these other great traditions amongst the employees, um, that are, ah, you know, both intellectual pursuits, but, um, but bringing out the best in competition.

Speaker A: Very cool. What's a technology trend in AI that you're most excited about right now?

Speaker C: I think like, you know, like openclaw, like the, the ability to like kind of democratize access to agents and let people really play with anything they want to automate. Um, so much of our lives revolve around our computers and there's so many opportunities for automation and letting just people play. Um, uh, again, going back to my origin story, we had TI81s and then, um, uh, a couple of us in high school made video games that you could play across, uh, the data link cable, like a scorched earth and things like that. So when you give people access to that technology and you democratize that access, a lot of cool things can happen.

Speaker A: I didn't know if you're going to say drug wars. That was a big one on the TI83. Um, all right, so, uh, who's somebody in the defense tech ecosystem that we should have on this podcast? Who's someone that you would, uh, recommend?

Speaker C: Um, so I was just at joint staff cdao.

Speaker B: Uh,

Speaker C: uh, Harley Stout, who's the uh, CDAO for Joint Staff, Uh, and it's a special staff section. He's ex booze. Really Interesting. Right? Like, um, those guys are at the pointy head of sphere, um, working both looking at all of industries, defense tech. And how did they employ it? Um, you know, for operational use. Um, so he's interesting. Um, Bianca Hurley, who used to work for me, she's the, um, uh, chief AI Officer for joint staff. Um, you know, those are the ones that come, come off the top of my head, but there are so many others at OSD that would be, you know, worthy to be here.

Speaker A: Nice. Yeah, we'll be, we'll, uh, be tapping them shortly. Always love a good, uh, you know, guest referral. Referral, yeah. Uh, so you're a serious squash player. Um, what's one lesson from squash that actually applies to building companies?

Speaker C: I think it's consistency, like Tom Brady said it best. And I'm of course, the season ticket pats, um, I'm a pat season ticket holder. Is to be successful in life, you don't have to be extraordinary at all, right? It's consistency and it's the thing that others don't do, right? Being consistent day in, day out. Uh, and, uh, I find that true in business that the most remarkable business people that I've worked with are just consistent. They put in m that effort day in, day out. Uh, they don't take any rest days. Um, and I see a lot of extraordinary people. Um, I went to a high school where I think it was the number one high school in the nation at the time. Still might be in the top four or five. Um, sorry, we just went over my one, uh, hour, uh, uh, timeline. So it's just reminding me. Um, but consistency, right? Um, it's true in everything facet of life. Very true in squash and very true in business.

Speaker A: Very cool. Nice. Um, so you balance being a founder, a professor, Cornell family man, you kind of started here with holding a 1. What's your secret weapon for managing time?

Speaker C: Uh, not sleeping. I think, uh, you know this as a recent father, uh, sleep is at a premium. Uh, I don't do a lot of caffeine either, so, uh, I drink a lot. Like my go to snack in the middle of the night as you get older too, right? You just put on the pounds is sparkling water, and if you drink enough of it, it'll keep you up or it will, uh, make you fear wetting yourself. So that's my secret weapon. Yeah.

Speaker A: Topo Chico is my weapon too.

Speaker C: It's so good. The mouthfeel is the Glass bottle. If you have a, um, podcast on the mouthfeel for the varied sparkling, uh, waters, one of my good friends during quarantine, you get really bored. In my quarantine group, we were talking about the mouthfeel of different sparkling waters. Toby Chico came up on top.

Speaker A: Oh, uh, that's awesome. Getting nerd out on that.

Speaker C: Exactly.

Speaker A: All right, uh, what was your very first car?

Speaker C: Toyota Tercel. Uh, it was a four speed. It was manual. Um, I had saved up all summer long to buy it. My parents put in the difference of money. It was from Woodbridge, Virginia, and the dealership is the bottom of a hill. Uh, they didn't realize that it was a manual. I didn't realize what a manual car was. Uh, and so they went to drive home and I learned, uh, via the sales guy, how to get up a hill using the clutch, uh, on a four speed. Uh, and I learned very quickly how to drive that manual car.

Speaker A: Yeah, pretty sure nobody knows how to drive a manual these days anymore, so.

Speaker C: Oh, definitely.

Speaker A: Good skill. Good skill to have. Uh, and then a closer. Uh, yeah, a charity or corporate philanthropy that's near and dear to you.

Speaker C: Um, because I do a lot with squash is the urban squash program. Um, they have, uh, deployments all, all across the US and it's a great charity. Um, uh, urban kids come in, they learn how to play squash, underprivileged urban kids, uh, and then they also pair with tutors and other things. They've been great about getting these kids into college, playing on teams. Um, squash is a relatively privileged activity. So as those kids go out to maybe potentially work at Wall street or work for these founders. The founder of whoop, um, is a billion dollar unicorn. He's ex squash player from the Harvard squash team. There are a lot of us around there. Brian Roberts from Comcast. Um, so, uh, it provides all these opportunities beyond just sport. They pair sport with education, with a pathway and license. I love that charity.

Speaker A: That's very cool. Sounds like a valuable network to, uh, participate with and learn a little squash skills along the way. Uh, that's a wrap, guys. A beer. Darren, thank you so much for spending time with us and talking through, scaling, deploying AI across government at large. So thanks for joining us on the pod.

Speaker C: Thank you for having us.

Speaker D: Great to be here. Thanks, guys.

Speaker C: Sam.