Log in to watch

Log in or create a free account to watch this video.

Log in
AI Summit Bletchley Park 2026
Share
Download slides

Mission Critical AI: Achieving Mission Speed Through AI-Enabled Rigor

Jeremy KnuppJK
VP, Global Command and Control Solutions, Northrop Grumman

JD Black and Jeremy Knupp of Northrop Grumman show how AI is speeding up mission-critical defense engineering without giving up rigor. Jeremy describes the drone threat facing air and missile defense and a rapid prototype built on the AiON product, and JD explains why agentic workflows need deterministic, measurable processes for requirements, interfaces and verification, asking what changes when rigor itself becomes free.

Chapters

Full transcript

The complete talk, organized by section.

Host Intro (Gene Kim)

The next speakers include JD Black. He is Chief Engineer at Northrop Grumman. Northrop was founded in 1939 and currently has over 43,000 engineers building some of the most advanced aerospace and defense systems in the world, including the amazing James Webb Telescope.

I met JD in 2017 because of his heroic efforts to bring modern software engineering practices across his industry. In 2023, he presented on how to integrate systems engineering, one of the last bastions of people to integrate into the technology value stream, showing how we can integrate version control, automation, and fast feedback in some of the most demanding environments.

So since then, in 2017, JD and I have been on so many adventures together. But one of the memories for me was when he told me how much fun he had spending weekends using AI on his projects.

So here is his mission: that when people go to work on Monday, his fellow employees working on some of the most consequential missions there are, that they can use the same tools that they use at home, at work.

And so he will be presenting with Jeremy Knupp, Vice President of Global Command and Control Solutions, who I got to meet many years ago, and I'm so happy that he's in his new role. Here's JD and Jeremy.

Jondavid “JD” Black

Well, thank you, Gene. It is absolutely fantastic to be here. I am a card-carrying computer scientist, so being at Bletchley Park has this special tingle about it, at least for me.

I actually started my career working the UK census—a very interesting and fun story, if anybody cares to hear about that—and then really went into, well, the last 20 years, basically working the “Star Wars” program. That's what most people recognize it as, before kind of moving to the dark side, to corporate, and becoming the chief engineer for our CIDO organization.

This is Jeremy.

Jeremy Knupp

Yeah. Thanks, JD. Jeremy Knupp. I have a less interesting history than JD, but started out in the US Air Force as an intelligence analyst and then moved into the defense contracting arena, really developing, as an engineer, systems that either collected, exploited, or disseminated intelligence data.

And then I decided I was having too much fun, and I moved over to running part of the business and have spent the last 15 years with Northrop bouncing back and forth between leading engineering and running parts of the business.

The last two roles were, before this one, leading digital transformation for the corporation, which is where I got to meet JD and Gene. And now, under Global Command and Control Solutions, the big part of my portfolio is integrated air and missile defense, which is very apropos right now with what's going on in the world.

So I get the part of setting the stage for the problem set, and then JD is going to talk about what engineering's going to do about the problem set.

But first, real quick, a little bit about Northrop Grumman. Large, top-five defense contractor. We have about 100,000 employees. As Gene mentioned, most of them are engineers, so I consider us to be an engineering firm. However, we have a lot of manufacturing space to produce the things that the engineers design.

Talk about our capabilities. We build capabilities from the depths of the ocean to the vastness of space, and I'll touch on the pictures, because that's probably the most interesting part of this slide.

Upper left is the B-21 stealth bomber. Upper right is an artist's rendering of Golden Dome, probably a modern sim of Golden Dome. Bottom left is the Artemis rocket boosters, which are the largest rocket boosters in the US inventory. To the right of that is the James Webb Space Telescope that Gene mentioned. To the right of that is the Manta Ray, which is an autonomous underwater vehicle. And then to the right of that is our SABR radar, which stands for the Scalable Agile Beam Radar, which is the main combat sensor for the F-16 fighter.

So my problem set for my business is integrated air and missile defense, and I'm going to cheat a little bit and use a video here to set the stage.

Maybe one more.

A little bit of marketing in there. So that sets the stage, but I don't even think it fully represents the complexity of what we're seeing. With what's going on with the conflict in Ukraine, we probably haven't seen transformation in contact since World War II like that.

With the mass mixed raids of 800 to 1,000 threats, think about the cognitive overload that does for the operators, the defenders. And we've seen the US Department of War, a year ago, talk about they wanted to implement AI and automation with a human on the loop, and they've shifted that narrative. They've now come to realize, especially with Epic Fury in the Middle East, they need full automation. It's too much, too fast for any operator human in the loop.

And then the mixture. So all the way down to mortar, artillery, rocket, the five classes of UAV threats, up to cruise missiles, tactical ballistic missiles, longer-range ballistic missiles, and now hypersonics. So just the mass and the complexity.

And now on top of that, we're seeing our adversaries innovate almost on a nightly basis. So changing the flight profiles of the threats, the behaviors, the characteristics. Think about what that does to our algorithms and our AI. What if the duck doesn't quack and it barks? Is it a duck? Will the system know that it's a threat?

And then the change of that behavior on our systems, where we try to intercept the threat at the optimal point in the profile. If the flight profile is changing, where is the optimal point to intercept?

So we have been thinking about how do we out-innovate and do it in a more rapid fashion than our adversaries. And I have an exemplar here of where we're using mature technology, mature AI, so think TRL level eight or above. Do rapid innovation integration so that we have a new capability that's greater than the sum of its parts.

We have a product called ION. It's AI on net. It's a counter-UAS product. And what it does is it integrates multiple sensors, whether they be RF, optical, or acoustic, to identify a threat using all that information, fuse a track, and then integrate with a bunch of kinetic and non-kinetic effectors. So the system will decide, “Yes, that's a threat. There it is, and I'm going to use this effector against that.”

However, we need a ride into the fight. We found this small company out of Phoenix, Arizona, called Swarmbotics AI. They create small UGV Fire Ants about this large, and the US Army right now is using it to hang it off the sides of tanks. And when they detect armored vehicle threats, they pull the UGV off, put it on the ground, put a C4 payload on it, and the AI takes it in underneath the other armored vehicle and detonates, disabling the adversary's tanks.

So we were thinking, “What if we could use that platform in a different way?” So myself and two other people from Northrop flew down to Phoenix, met with their CEO, CTO, and their lead software engineer, got really excited, and in the course of a couple of hours in an afternoon, sketched out an architecture where we would integrate our ION with their UGVs, and then challenged ourselves that in six weeks—so we went there in mid-October—we're going to do a demonstration of the new capability in mid-November.

And what you're seeing on the right there is an overhead view of my office in Huntsville, Alabama. There's a field next to my office, and the demonstration's going to be using our ION product to geofence a non-drone zone, so that's what the NDZ is. Create defense planning, so optimal placement of the four Fire Ants that we're going to use for this demo, and then task them to go out and protect this non-drone zone.

Three of the Fire Ants are going to have acoustic sensors, and then one of them's going to have an electronic warfare effector on it. So the only human in the loop in this entire demo is we're going to fly a drone into the NDZ. We're going to use the three acoustic sensors to detect the threat, use ION to triangulate where that threat is, and then task the EW effector to bring down the threat.

And then one of the Fire Ants, using one of the optical cameras on the Fire Ants, is going to go out to the last known position of where that drone was and then confirm, yes, it is a drone, and yes, we did bring it down.

So that's an example of how we're thinking about rapid innovation with AI, at least from a mature AI. What JD is going to talk about is new capability and rapid development of new capability.

JD.

Jondavid “JD” Black

Thank you, Jeremy. I like to frame this, and Jeremy's discussion did a good job of that. We really have two fundamental constraints that we deal with. We've got physics and we've got economics.

So much of what we do is driven by the physics of what happens in the world around us, and it's a true constraint. But we also have to comprehend the economics of it. In my previous life, it was the missions that we fought. We would be shooting down $5 to $10 million threats coming at us with $100 million bullets. The economics don't work out well.

The same's true when we look at building new systems. We're going through a revolution that we see right now. And as we heard from Fernando, we can't put the armies of people, the years of time into solving these problems. We have to move faster than that.

And so where we're seeing this happen is really building on what we've learned in the last 20 years, where we saw product delivery really transform with DevOps. And that focused on what happened from that commit to version control out to delivery into operations. We're seeing the change now, shifting how we go from idea, concept into actual capability that can be pushed through that process as quickly and as economically as possible.

We've lived through a hype cycle. Who recalls when we were all getting fired and being replaced by prompt engineers?

We've seen it progress over time, and at least for me, it's always been really interesting. I always learn something new, but I always am left wanting a little bit more. The next thing comes, and the next thing comes, and we're continuously improving, but we're not quite at that point where we have what I would call behavioral determinism that we can rely on to do complex, important work.

I kind of call it the Anchorman problem: 60% of the time, it works every time. And when you do the kind of missions that we do, that's not good enough. We need to see something substantially better than that.

And so we go back to building on what we already know. And again, over the last 20 years, we learned building out the people, processes, and tools and really optimizing that allows you to maximize the effectivity and the performance of your organization.

As we bring the agentic AI technology to bear, we're seeing a parallel. It's really about agents, workflows, and skills. Thankfully, agents and skills are largely being standardized. You can build your Markdown files, or your vendors are selling them to you, so that you can get the kind of instruction that you want.

The piece that we found to be missing is that workflow. How do I know that the agent didn't skip three steps along the way? I see it all the time. We have standard processes that we follow, and we have to follow for regulatory or legal reasons. And when step five gets skipped, it's a really bad day.

So we're working to put that predictable atomic behavior into the constructs that allow us to rely more on these technologies. And we, again, can lean on what we've learned. Value stream management and understanding how to get your flows correct is still incredibly important. Configuration management of those structures so that you can deploy them and repeat them is really important. And then good testing and validation is also critically important.

The talk this morning Jacqueline gave really resonated with me. We are working and have built workflows to essentially implement our system development life cycle, and now we can do pieces of that pretty reliably and deterministically. The challenge is becoming more and more, not “Can you do the thing?” You almost certainly can. It's “Should you do the thing?”

Do you have an executable strategy that's actually driving you to accomplish something in your strategic vision? And we're really starting. We've focused so far, as I tell my team, we're living in the box of outer loop and inner loop right now. And in the next year, we're really going to start pushing hard into the strategic vision, working with our business management organization, our CFO's office, to connect those dots and know that we're moving the needle for the business, for our customers, and for our missions in meaningful ways with every step.

One of the most important things that I think is coming to bear right now is the systems engineering piece of this. We talk a lot about code and how important pushing code out is, and that's absolutely true. But do we understand the systems?

At Charlotte, we saw a lot of really good talks, and one of the consistent themes was we don't know how to manage our requirements. We don't have our stuff written down. It happens all the time.

We've been kind of the old dodgy contractors who live and breathe requirements and verification and interface definition and system decomposition. Well now, as of last year, OMG has published SysML v2, which is a code-based way to do model-based systems engineering that you can capture all this. And we're building out our agents with the right instructions, with the right templates to actually automate this process just as much as you're automating your code development.

It allows us to build that end-to-end capability: concept definition, requirements, interface, all the way through verification.

And that looks like this. We're really starting to take it two different ways. Number one, build out the workflows, understand the steps in your sequence in a way that your agents can deterministically and measurably execute that. You want objective evidence that every step along the way was done. You want to know how long it took, how many tokens it took, how many loops you had to go through to get there.

And you can see we're doing that through enterprise AI with that system development life cycle, and that's a snapshot of what it was about four weeks ago. We're continuously optimizing and improving that, and working with Jeremy and his team, we're really starting to lean in on the mission AI.

How do we take the same fundamental building blocks, the same need to solve complex problems, and put the technology to bear in ways that can give us behaviorally deterministic outcomes? We know we didn't skip important steps along the way to make poor decisions in an often safety-critical and mission-critical environment.

So with that, what are we looking for? What are we asking for? A couple of things.

The work that's going on right now in China is amazing, but we need good Western models that we can rely on. If you support or work with organizations doing that, please continue to do that. We have a lot of need right there.

As we see the technology and the models themselves advancing, the tooling around them is super important. We need to get really good at version control of our agents and dependency management associated with that.

Restarting my harness is not an implicit approval to change the brain of my system, and it's really frustrating to see some of the things happening. I know we're moving fast. Fast doesn't mean we operate without the appropriate controls and guardrails. And so as we think about how to build these things, let's think about controlling what we have in the right ways.

Agentic workflows. We're seeing this—I guess it was last week, we saw GitHub put out workflows as a part of their harness. I'm starting to see workflows become a thing. If anybody would like to partner with us to evolve how we can standardize the definition of repeatable workflows in the market, I would love to work with you on that. We've had some great success so far with that, and just the ability to exquisitely measure what you're doing with your agents is, we found, to be incredibly valuable.

Forums like this, please, they help us so much. I hope this helps you. But let's just continue to share.

And then I'll ask if anybody has any thoughts on it. We talk so much about what do you do when code is free, but now we can step back and put the right controls in place. What do you do when rigor in your systems and your processes becomes free? How do you apply that? How does it change how you deliver value into your space?

It certainly is going to change how we deliver value into our market.

Thank you.

Thank you.

One membership. The IT Revolution library.

IT Revolution's books, the video library, the courses, and the MCP. The papers stay free.

Coming January 2027