Smack is Building the First Defense AI Lab for Decision Dominance ⚔️

CV Deep Dive

Today, we’re talking with Andy Markoff, Co-Founder and CEO of Smack Technologies. 

Smack is a frontier AI lab building purpose-built models for national security, focused on delivering “decision dominance” across military mission sets. Founded by Andy and his co-founder Clint Alanis (both former United States Marine Corps Forces Special Operations Command or, MARSOC) after firsthand experience with the friction and fragility of modern targeting workflows, Smack is designed to fuse and analyze multimodal data streams in real time, then translate that understanding into actionable options for commanders: faster than adversaries and optimized for campaign-scale operations. Under the hood, Smack leverages deep reinforcement learning and a proprietary training environment that combines the physics of peer conflict with encoded human tactical and operational expertise.  

Today, Smack is actively working with the Marine Corps, the Navy, and the Air Force: supporting different decision horizons from deliberate planning to execution cycles to real-time dynamic operations - and is pulling those efforts into a unified workflow that connects planning, execution, and live adaptation. The company recently closed its Series A and is expanding across the Joint Force and internationally, while continuing to hire top AI research and engineering talent.

In this conversation, Andy shares how Smack was born out of operational pain points from Mosul to the broader military, why reinforcement learning and high-fidelity training environments are central to building effective warfighting AI, and his vision for how specialized frontier models can reshape decision-making in conflict.

Let’s dive in ⚡️

Read time: 8 mins

Our Chat with Andy 💬

Andy, welcome to Cerebral Valley! First off, introduce yourself and give us a bit of background on you and Smack. You have a very interesting background, tell us about that.

My name is Andy Markoff. I’m the co-founder and CEO of Smack Technologies. Before Smack, I was a member of a cult known as the United States Marine Corps for 10 and a half years. I spent half my time in the infantry and the other half as a Marine Raider at MARSOC, doing the usual deployments to Iraq and Afghanistan.

For me, the genesis of Smack goes back to 2016. I was a special operations officer for the task force running the Mosul operation. I ran the kill chain for the Kurds and the Iraqis in Northern Iraq with our teams embedded with them. We ran the targeting process in PowerPoint and the broader Office 365 suite. We would weaponize targets on whiteboards, build Gantt charts in PowerPoint, crash SharePoint every morning trying to upload them, and it was a mess. I’d say I left the Marines frustrated with a lot of the processes around how we ran the targeting cycle. I went to work at Palantir for a bit, trying to figure out what I wanted to do when I grew up. I did a couple other jobs in the tech industry, tried being a mountain guide for a time, and realized I was too old to haul people around in the snow and probably needed to get a real job. 

Then my co-founder, Clint Alanis, and I started talking. We served in MARSOC together, and he was retiring in 2023. He said nothing had changed in the last 10 years about how the kill chain was being run, and we should work on it, because it really isn’t going to work when you think about China or Russia. He was someone I’d always wanted to work with again, so the opportunity made sense, and that’s how we ended up here.

How would you describe Smack and its mission to the uninitiated developer or researcher?

At the end of the day, Smack’s mission is delivering decision dominance to the Department of War, and to U.S. allies and partners. So you might say, what is Decision Dominance? The simple answer is, it’s the process of taking multimodal data streams, analyzing them in real time, and then converting that analysis into the right decisions across the kill chain faster than our adversaries. 

So now you might ask more questions: How do we do it faster than our adversaries? How do we do it in a way that is optimized for a campaign writ large, not just a small unit in a very small place? What do you think about a theater’s equities when you do that?

That gets to what Smack is: the first frontier AI lab building for national security. We are building domain-specific frontier models, not general-purpose models trying to figure out how to fill your Instacart order and run military operations. They are built for a single purpose: deterring and, if required, winning World War III. They have a heavy focus on a very detailed training environment, leveraging reinforcement learning and deep reinforcement learning to build those models and deliver them to the warfighters.

Defense companies are often document-driven rather than requirements-driven. How did your time in service influence what you’re building with Smack?

Clint and I do not have to guess what end users need. The reality is we were the end users, and both of us were instructors who taught many of the current end users. I was an instructor at the Marine Corps Top Gun in Yuma, where I taught fires integration. Between Clint, me, and a couple other members of our team, we have unique, deep expertise in this problem set, and we fully understand what’s broken. I experienced it over the course of my career for 10 and a half years. Clint experienced it for 20 years. Ray Gerber, one of our business development leads, was Admiral Paparo and Admiral Aquilino’s fires planner for the last several years of his career, looking at this problem set in the most important theater in its entirety. I’d say we really understand what is needed, and we bring a unique perspective.

On the one hand, I keep joking we have to stop hiring Marines because we have too many crayon eaters in the club. But the interesting thing about the Marine Corps is it is full of generalists. If you look at the Joint Force, you will find Marines at the seams between the Air Force, the Navy, and the Army in a lot of the fires cells. The reason is the Marine Corps is a mini joint force where we understand how to integrate a wide range of capabilities. It is fundamental to how the Marine Corps trains its officers. You see that getting farmed out across the Joint Force. 

I say that because one of the biggest issues is our acquisitions program to date has been very service-based. The Navy has its own acquisition system, the Air Force has theirs, and the Army has theirs. People are trying to buy command and control capabilities, but they are buying them with the service interest, not the Joint Force interest. Our team understands how joint doctrine comes together, where you can make artful trade-offs and say: “The Air Force does it this way, and the Navy does it this way, but we are going to make a design choice to do it this way because it is the best mediation of doctrine between two services that don’t actually reconcile.” I’d say we have the credibility to do that.

Tell us about your key users today to the extent at which you can. Who would you say is finding the most value in what you're building with Smack?

Today, we are actively working with the Marine Corps, the Navy, and the Air Force. There are other contracts coming, and it’s highly likely that we will deploy across the Joint Force later this year.

When we think about our product and who is using what, we have the model and we have applications. A lot of the difference between a product and an application that leverages the same model is the decision horizon the user is working in. 

The military basically has three decision horizons: what are we doing for the next one to six months, which is the deliberate horizon; what do we do for the next one to four days within the context of the operation or plan we built in that deliberate cycle, which is the execution horizon; and the dynamic horizon, literally right now, what are we doing. With the Marine Corps, we are working on the deliberate horizon, what are we doing for the next one to three months. With the Air Force, we are working on the execution horizon, what are we doing for the next one to four days within the context of this plan. With the Navy, we are working on the dynamic horizon, literally right now, what are we doing. Those are all part of the stack and part of the product suite. 

We initially prototyped different decision horizons with different services, and now we are pulling all of those together into a master workflow: here’s the plan, here’s the next execution cycle, and here’s what we are doing right now against that execution cycle, seamlessly.

Your website features the headline “AI for Decision Dominance”. What does that look like on the platform? How does this appear in your suite of products?

I’ll talk a bit about how we build the product first, and then how it comes together at a high level. Fundamentally, we’re using deep reinforcement learning. The training environment we’ve created adheres to the physics of a modern peer-level conflict, but not just the physics. It also captures the tactical and operational considerations that can only come from human domain experts. A lot of the secret is finding the right domain experts, extracting their knowledge, and encoding it into the environment. It’s not just that a jet of a given type can fly a certain way. There are tactical considerations about how they fly and why they do certain things. If you modeled the physics at an intractably granular level, you might arrive at many of the tactics, but to solve this problem in our lifetimes and get a better starting point, we encode human expertise and art into the environment alongside the physics. That environment is really our advantage, and then we use modern reinforcement learning algorithms to build the model against it.

If you think about the model’s capabilities, there are two. First is an orient capability: how we fuse and analyze multimodal data streams in real time. Think petabytes of sensor data. How do we geospatially align it over the earth and in time, then analyze it to understand the current state of the environment?

Second is: given the current state, what do we do about it? Without going into details, if you think about strike waiver or pulse planning in a peer-level conflict, traditionally the U.S. would go in, achieve air superiority, and then do whatever we want. Against peer adversaries, we can’t assume that. Every time you want to break in, you have to create a bubble: break in, do what you need to do, get out, and assume everything collapses. 

Now imagine a 100 million square mile theater, pulling sensors, shooters, and logistics tails from all over the world, coordinating the timing of individual munitions hitting targets to plus or minus 10 seconds, and doing that over and over again. Today, planning those pulses takes three to four days. You need to do it in minutes to seconds, and then re-update it based on what’s actually happening in real time, in seconds. A lot of the use cases we’re focused on are helping users decide what to target given a network system and goals, identify the right nodes to go after, determine the next pulses of assets to send against those nodes, and then adapt when something changes in execution: an unexpected target pops up, weather changes, or an asset has a sensor malfunction and has to go home. You need to fix it without shifting time on target, or if you do shift time on target, understand what that means for all the assets coming in from everywhere. 

The goal is to create options for a commander to choose from and then push those into execution.

Some may say 2025 was the year of reinforcement learning. From our past conversations, yourself and your team have a huge focus on RL, which for warfighting, makes perfect sense. How are you using it today?

It is a huge elephant to take bites out of. I’d say part A is that we know the right places to start. If you try to build the environment for the entirety of the action space for World War 3, it’s intractable, and you don’t even know where to begin. Where we’ve focused is: what are our contracts, what is the problem they need to solve, and what is a snapshot of the overall environment that we can get moving on? Part of the secret we’ve had is right scoping: which part is the most valuable to start?

The other piece, which I was not expecting two years ago but has become the reality, is that there is an internal product we’ve had to build so we can iterate on the environment. It’s essentially a user interface where I can take a network of domain experts, some on staff and some we use periodically, interview them, and turn that into key proprietary data. There are different ways we store and use that data. You can RAG-extract it into an agent’s personality in a sim, or encode it as constraints on how the environment is allowed to behave. 

There’s an art to how we apply that information. The way I like to think of it is: we make movies about World War 3 for the RL agents to watch and learn how to fight World War 3. Then domain experts like me can watch the movie frame by frame and provide feedback: this would never happen, and here’s why. We annotate the frames, capture what’s wrong, and go back to generate a better seed scenario. Once you have one really good seed scenario, you can riff thousands and thousands of variations on it. Then the next piece of work is: what’s the next right seed scenario? 

To recap, we have unique insight into the right seed scenarios for the high-value problems. Per seed scenario, we interview domain experts, encode the physics, take a shot at how a scenario could play out, and then let domain experts review it frame by frame, provide feedback, and keep iterating until we get a good one, and then expand variations from parameters that are worth exploring.

There are many companies working on defense AI, but it seems like you’re the first pure AI lab in the space, cross-pollinating operational and research expertise. What else sets Smack apart from your competitors?

I’d say some of it comes back to my experience in the Marine Corps. You’re a jack of all trades, which used to drive me insane. You join thinking you’re going to do a certain thing, and the second you can avoid getting your platoon lost in Iraq, they hand you a mortar platoon and tell you to make sure the rounds don’t land in base housing during training. Once you can do that, they move you again and say, now you’re a Top Gun instructor, and you’re thinking, I don’t know anything about flying. They’re always broadening your perspective, but you end up understanding all these different cultures and capabilities that have to come together.

I’d say the unique perspective we have is that everyone tries to map us to a category: are we like Palantir, Anduril, or OpenAI? My answer is we’re none of them. We need scientists and we need engineers, and they’re different personalities. In many ways, engineers are similar to Marines: I have a due date, I’m going to work backwards and build toward it. Scientists are solving open-ended problems: I don’t know what the answer is, I need to run a bunch of experiments, and I don’t know when they’ll be done because we need to do it right. 

Those are different outlooks that are hard to bring together, and then you add military expertise on top. Even within the scientist side, we have a broader range of disciplines than most companies: physicists, operations research specialists, RL specialists, supervised learning specialists, and there are hardware aspects to some of the products we’re building that we’ll need to implement. The range of scientists is broader than most companies, the range of engineers we need is broader than most companies, and you have all the different service disciplines in growth. 

There’s a desire to push everyone into a single culture, but ultimately we want to win, and we want people who are mission-oriented around national security and deterring peer-level conflict. At the same time, we make room for a wider range of personalities, beliefs, and ways of doing business because we have to. We need the whole team to work together, even though they’re very different. Having seen that in the Marine Corps, where my job was to take different Marine disciplines, joint force disciplines, and partner force disciplines and get them all aligned even though they don’t see the world the same way, I’d say that perspective and the willingness to culturally allow for it is one of the things that makes us different.

What has been the hardest technical challenge around building Smack’s product suite?

That’s the balance between the physics models and the human priors. We have a physicist and a physics lab, and he has a deep background in this. Our physics models are deterministic, but as you start thinking about sensor collection and the impact of the environment on sensor collection, it becomes computationally intractable to get super granular with a deterministic physics model. A lot of the difficulty has been deciding what tradeoffs to make on the deterministic physics, and how to make up for that with good human priors.

It’s funny, because in many ways I empathize with physicists. Tactics are, in many ways, a function of trying to observe and shoot first, and be observed last. If you think about the whole OODA loop, it’s pretty generalizable. In theory, if you modeled the capabilities of all the platforms, you would arrive at the human priors we use today just by getting the physics right, but that would take so long it’s computationally impossible. So a lot of the internal arguments are about where we make the tradeoff, what is worth exploring because the humans might have it wrong, and what the humans have right and we just need to start there, rather than going down a rabbit hole trying to model the physics perfectly. That messiness, plus building the infrastructure to get the environment right, has definitely been painful.

You just mentioned you have a team of physicists, and earlier you mentioned ex-marines and researchers. Could you talk a little bit more about the team composition at Smack? 

At a high level, there are two major parts of the organization: the growth and mission operations side, and the product and technology side. I’ll talk about the easier side first, which is growth. That’s where a lot of our Marine veterans with the right network and domain experience sit. They help us go out, understand the problem and the end users, win contracts, and successfully deliver on them. They also play a deeper role on product than you’d see at many other defense tech companies. Because the environment matters so much, and because recent domain experience matters so much, the growth team interfaces with the product team regularly. They’re often the first people we interview. If you think about a proprietary Smack dataset, we’re pulling knowledge out of the brains of the growth team. Over time, that knowledge decays. 

I’ve been out since 2017, and the relevancy of my knowledge decays, and the same is true for all of us. When we hire more mission operations and growth people, we bring them on, pull their knowledge out, and over time they become savvy around the product. They understand the sales cycle and they’re incredibly useful, but we always need that inflow of domain experts who know the end user, help us win contracts, and help us build the product. Growth is about winning contracts, and mission operations are about successfully delivering the contracts we’ve won. The reality is we’re a small company, so we hire people who end up doing both right now, but in time those will split.

On the technical side, we’ve recently reorganized. It’s become apparent that fundamentally we’re building frontier models because there’s a requirement and a gap in the models that are out there. 

We’re organizing more like an AI lab: a research group, an applied engineering group, and a deployed engineering group. The research group builds the core technology, including the environment and the models, and does the tuning. In that capacity, the growth team and our domain expert network interface with research on fine-tuning the environment and the model. The applied engineering group takes the model and builds applications that let users leverage it for different workflows. Those workflows are warfighting functions: joint fires, logistics, force protection, intelligence, and information. They’re building applications that leverage our models toward specific workflows, with specific artifacts that have to come out the back end. That’s what turns the models into a usable product. 

Deployed engineering looks like what a lot of defense tech companies adopted from what Palantir did: deploying engineers forward and embedding them with users so they can iterate quickly on how well the product is actually working. Today, because we’re relatively small, applied and deployed are effectively the same team, and in time those will split. I’d say people cross-pollinate between groups, with research spending time in applied and applied spending time deployed, but functionally those are the three main functions.

Your team is highly diverse in background and skillset. When you’re bringing new folks onto the team, what are some common qualities you look for across the entire organization?

At the end of the day, we look for people who want a mission, not a job. This is definitely not 9-to-5 work. Especially as a startup, it’s full-on all the time, and there isn’t a whole lot of balance. I’d say it’s important that if you’re going to spend all this time doing it, you like what you’re doing, and not just like it, but believe in it. 

We want someone who wants the mission, and at least at this point in their life, that mission is national security. That’s what we’re doing, and we don’t beat around the bush. No one wants war, and if you want that, you’ve never experienced it. But the reality is that if it happens, we see ourselves as a company founded by warriors. This is our profession. For Clint and me, our entire lives are oriented around this. We’re building software that helps figure out, if we go to war, how do we crush our adversary. It’s important that people are okay with that. If you’re not, that’s okay, it just isn’t the right place.

Second, we’re here to win. Not everyone gets a trophy, and winning matters. I’d say a lot of the cultural values people try to get at are downstream of winning. If you focus on winning, a lot of the other stuff falls into place. 

Third, we’re a culture of rejecting the iron triangle. The Department of War doesn’t care that you think you can only have two of the three. They need it yesterday, they want it for less money, it has to work, and you need to figure it out. No one cares that it’s hard. As the Marines used to say, even the laws of physics can be bent with a little motivation. We’re going after every contract related to decisions, and it needs to be better. I can’t sell the government's worst software just because we’re a startup. It has to be good, it has to be cost-effective, and they needed it in the stack already, so get moving. That’s not an environment for everyone, but it’s one of the distinguishing features we look for: people who can work in that type of environment.

Does that law also apply to your physics team?

Laughs They do. Greg [Greg Passmore, Chief Scientist at Smack] fights with me about this all the time, but he’s very motivated to overcome the laws. Greg would say to find new laws of physics first, then overcome them.

Lastly, what’s your personal favorite part about building Smack?

One of the hardest things about getting out of the military, for a lot of veterans, is that you lose your sense of purpose. For ten and a half years, I knew what my purpose was: defending our country. It’s given to you, so you don’t have to think about it. You just have to try and do it well. Then you get out and you don’t have that, and you’re left asking, what do I do with my life, and what feels meaningful? 

For Clint and me, I’d say there wasn’t another thing either of us were born to do. It felt like it was time to move on from literally being in the military, but there isn’t another thing I identify as other than being a warrior, and Clint is the same way. This has been an opportunity to take the knowledge Clint and I have, fix the things about the military that we knew were broken and know are still broken, and really give back while still serving our country in the national security space. I’d say that has been really meaningful, and it’s been a hard journey since getting out to find my way back into that.

The second piece is that I get to work with people I want to work with, who are awesome. Clint was one of the people I always wanted to work with again, and the same goes for a new executive we’re bringing on. It’s like all of the best people from the military, who were outstanding at their jobs and who Clint and I both worked with, now work at Smack. On the engineering team, we’ve brought in world-class talent, people who perform, who care, and who have the same mentality, but in a technical discipline I didn’t experience in the Marine Corps. 

Having this high-performing, elite team working on a really hard problem with crazy deadlines and obstacles, and all trying to figure out how to solve it together, is what I felt like I was missing after I got out, and now I’ve found a way back into it.

Anything else you'd like our readers to know about the work you’re doing at Smack?

We just closed our Series A, we’re actively expanding across all six services and internationally, and we’re hiring top AI talent. There are a bunch of places where AI engineers and AI researchers could work, and there are certainly places that pay more cash than Smack could pay. But what I’d say is this:

If you’ve ever wanted to serve in the national security space, or in the U.S. military, and you haven’t had the opportunity, we’re offering a chance to work on this problem set at the deepest level in a way that will have a real and lasting impact on our users. A small, fast-moving elite team that’s super focused: this is the place for you. That’s the environment we want to create, and in many ways the value proposition of working at Smack is working on the hardest AI problems. You’re going to work on the cutting edge of deep RL, and you’re going to do it for a problem set that’s existential.

Stay up to date on the latest with Smack, follow them here.

Read our past few Deep Dives below:

• OpenAI’s Codex Lets One Engineer Ship Like a Team of Ten

• InterviewGuard is Making Interview Fraud Impossible

• Augment Code is Your AI Coding Agent for Real-World Codebases

• Nebius is Your Full-Stack AI Cloud and Inference Platform

• Daytona - Composable Computers for AI Agents

If you would like us to ‘Deep Dive’ a founder, team or product launch, please reply to this email ([email protected]) or DM us on X or LinkedIn.

CV • Instagram • X • All Events