Pipeline as a Product: How Project Linchpin Plans to Deliver Artificial Intelligence for the Army

Pipeline as a Product: How Project Linchpin Plans to Deliver Artificial Intelligence for the Army

The U.S. Army knows that artificial intelligence (AI) and machine learning are critical to maintaining an edge over current and future threats in data-driven operations. AI-enabled systems will be needed to support a broad range of tasks rapidly and accurately, and on the modern battlefields in Ukraine and in any future conflict, success comes from making decisions faster than the adversary — in some cases, autonomously. AI and machine learning can also assist with everything from predictive maintenance to target identification. But despite these critical capabilities, only recently did the Army begin to ask an important question: “How do you deliver secure, trusted AI to the systems that need it?”

The Army’s answer is Project Linchpin. Named for the role that the Army hopes to serve in future conflicts — a “linchpin” in the Indo-Pacific theatre, for example — this project takes the standard AI and machine learning operations pipeline from the technology industry and modifies it to perform in a secure government environment while protecting operational data. Project Linchpin offers a secure structure that could be replicated across the Army to deliver AI at scale. It also acts as a literal linchpin, a vital central component, setting technical and ethical operating standards and interfaces for these emerging capabilities. Project Linchpin puts the Army and the Department of Defense on a path toward the first program of record that will deliver AI and machine learning capabilities and propel AI beyond labs and experiments and into operations.  

A Pipeline 60 Years in the Making

The conversation about AI is not new because AI is not new. The phrase “artificial intelligence” was coined in 1956 and the Department of Defense has been funding research in the field since 1963. However, the modern conversation around AI is a result of both the increase in computing power and accessibility of data. The Department of Defense’s research and development under Project Maven made senior leaders aware of AI’s potential to help make better decisions faster while maintaining situational awareness and understanding of the operational environment. The department formed the first operational organization for AI — the Joint Artificial Intelligence Center — to accelerate adoption of the technology in 2016 and formalized the Chief Data and Artificial Intelligence Office in 2022. 

 

 

However, for all the commentary from military leaders, policymakers, and government officials about integrating AI into the Department of Defense to create decision advantage or enable autonomous systems, suggestions are often limited to operational discussions around tactics and ethics. The lack of technical details around implementation can largely be attributed to a simple fact: the math is hard. In the Department of Defense and elsewhere, true expertise in the AI space is rare.

It’s also easier to focus on the AI performing a task because it’s sexy. It mirrors human intelligence aided by high-performance computing. But implementation is critical. In most commercial applications the risks of error are just an inconvenience. If Google Maps recommends a faster route that takes you into a traffic jam you might simply be late for your meeting. However, if an intelligence sensor confuses a school bus for a tank — or fails to detect the tank at all — the results could be catastrophic. No matter how much talk there is about how game-changing AI will be in the future, failing to plan for AI’s “back end” infrastructure is planning to fail.  

Fortunately, the time and investment in pathfinders gives Project Linchpin a head start. Five years ago, Project Maven introduced AI to the Department of Defense. This first AI “project” is the common name for the department’s Algorithmic Warfare Cross-Functional Team. The problem was simple: There was far more intelligence data collected from all the military’s sensors than could possibly be processed by human analysts. The answer, in theory, was also simple: The department could use AI to assist and increase the speed and accuracy of decision making on the battlefield. In 2018, however, the hard part was implementing this kind of AI system. Figuring this out meant figuring out how to invest, manage, and deliver AI in the Department of Defense more broadly.  

Because of Project Maven, a growing number of soldiers now know how to describe their AI requirements and have project managers who understand how to deliver them. These developments come at a time when the Army is learning how to employ AI and when industry leaders are calling for disruptive change so the United States can be competitive against its adversaries.  

With Project Linchpin, the Army is taking an important step to address its management and synchronization of AI and machine learning. After years of research, experimentation, and demonstration, there is now a template that outlines the management and delivery of AI for an entire portfolio of similar programs, like intelligence sensors. This helps to avoid a scenario where the Army competes with itself internally for limited AI engineers or pays into redundant development environments. Instead, program offices can focus on their operational requirements, knowing that a pipeline is running in the background that guarantees the reliable delivery of AI.   

On Oil and Electricity: The Boring Parts of AI

To achieve the Army’s modernization and digital transformation goals, investments should be made in the underlying architectures and infrastructure components that are necessary to make AI work. A future with AI-enabled systems and mission command platforms will require the Army to manage and secure its data, maintain innovative development environments, and use acquisition approaches that provide consistent growth in technical capabilities with developing AI talent within the Army workforce.  

Project Linchpin tackles both the technical requirements to build and deploy secure, trusted AI and machine learning models as well as necessary programmatic details including contract strategies, cost models, and the cyber security accreditation needed sustain the continuous integration and continuous delivery of these capabilities. The Army’s first use case will be narrowly focused on intelligence sensors that have requirements to perform many tasks well-suited for AI capabilities. 

When realized, Project Linchpin will provide a government-owned version of the industry standard machine learning operations pipeline, which can dynamically respond to changes in the environment, provide security measures against adversary attacks, and retrain and optimize AI to keep up with the pace of modern operations. Several things occur inside the pipeline: model training, test and evaluation, verification and validation, deployment, monitoring, and retraining. This kind of process delivers AI that can perform tasks as intended, but the real value that Project Linchpin provides derives from its relationship with users and their feedback. In each of the steps across the pipeline, soldiers will have access to the AI’s status and performance like a car’s dashboard, with the ability to provide feedback that helps improve AI performance and refine requirements for new capabilities. 

Even when Project Linchpin is operational, the development of AI capabilities will still be critically dependent on the most important element of the process: data. For AI development to be successful — or at least efficient — developers must have access to data. If AI and machine learning capabilities are the new equivalent of electricity, then data is the new oil that fuels these power plants. Army leaders have said they don’t have a data problem, they have a data access and quality problem. Therefore, making data accessible is critical to kickstarting a developmental pipeline, and will remain a key dependency for the success of any machine learning operations pipeline.

If this sounds complex, that’s because it is. Compared to people who work with deterministic software, finding people with a background in algorithms, data structures, neural networks, and, more broadly, computer science is rare. AI development requires an artistic touch. However, human capital is just one obstacle to the Army’s adoption of AI. Security is another.

 

 

In many cases today, AI systems are vertically integrated. That means systems are built as a full stack: hardware, software, and AI. While that can create some efficiency in the private sector, in government it results in proprietary solutions that limit competition and, more importantly, limit trust and security. 

Project Linchpin offers an alternative. By decoupling AI from software, the development and delivery process for AI can be responsibly managed. AI developers, engineers, testers, and security teams can be integrated into workflows that are governed by Department of Defense standards for performance, security, and explainability. The end-to-end development and deployment process allows cyber security measures to be built in up front, while creating mechanisms to monitor, check, and troubleshoot activities, all without exposing government data or AI and machine learning attributes to adversaries.

When thinking about the Army of 2030, and the expectation that there will be human-machine teaming enabled by AI systems, the importance of investing in secure development and delivery practices like this cannot be understated. The risks associated with employing AI and machine learning capabilities in combat operations are unlike those in any other private or public sector enterprise. False positives and false negatives can lead to slower, less confident decisions when time is critical to make decisions faster than adversaries. In worst case scenarios, if the Army relies on AI-enabled systems to detect threat formations but those systems miss their targets, false positives or negatives could lead to catastrophic outcomes. It’s not hard to imagine scenarios where a false negative leads to an undetected enemy counterattack, or a false positive leads to serious civilian casualties. 

Project Linchpin will stand alone as a separate program of record, a government-owned machine learning operations pipeline, which will mitigate these risks by breaking up the vertical integration discussed earlier. Instead of expecting every system developer to impose the same strict security measures, or expose government training data to uncontrolled environments, Project Linchpin will implement universal security standards, performance parameters, and barriers to entry for contractors, sub-contractors, and the entire supply chain. 

Getting Army Programs Out of the AI Business

The Army’s current experiments and exercises in AI and machine learning, like Project Convergence, demonstrate the importance of these capabilities to achieving “decision dominance” — the speed and accuracy of decisions and their effects — in operations.    

However, these bespoke use cases reveal an underlying capability gap across doctrine, training, and materiel solutions that Project Linchpin hopes to remedy. Instead of expecting every program that incorporates AI or machine learning technology to maintain distinct, technically skilled AI teams with developers, engineers, testers, and maintainers, Project Linchpin centralizes activities under a consolidated office that maximizes the experts and employees available. The Army has a focus on People First, which aims to create a culture that emphasizes talent management. Project Linchpin’s framework not only relieves pressure on the intelligence system and sensor program offices that it will support, but also incentivizes development and employment of AI skills and teams. 

An independent machine learning operations program office like Project Linchpin can also consolidate the available AI workforce and use their skills and models across the department to avoid potentially detrimental internal competition. For example, object detection from imagery data for intelligence sensors might be needed by multiple programs. Instead of each team competing for the same few data scientists and AI and machine learning engineers, Project Linchpin can employ one team to fix the problem for the entire organization. Models are trained and optimized within the project’s pipeline and delivered to each program for specific use. Project Linchpin is also responsible for the services like training and testing, as well as the optimization, deployment, and monitoring of the models once delivered.

This centralization allows Army programs to remain focused on their already challenging operational requirements and allow Project Linchpin’s machine learning operations pipeline to meet their technical AI and machine learning development and maintenance needs. Of course, the Army isn’t limited to its internal organizations, headquarters, and personnel to deliver AI. In fact, Project Linchpin not only creates an entirely new opportunity to leverage tools and services across the Department of Defense, but also to maximize private sector talent and innovation.

Accelerating Across the “Valley of Death”

Just as important as the capabilities themselves is the ability to maintain a sustainable, affordable acquisition approach. Decoupling the procurement and maintenance of AI from that of the software and hardware of the systems employing them will be transformational for Army acquisition, its science and technology segment, and the private sector as a whole. Examples of new and unique contracting options are emerging as the Army realizes it can partner directly with AI developers. These trends will continue as the understanding of AI and machine learning development increases. For the Army to manage an AI-enabled force, it will need to up-skill its workforce, increase digital literacy, and be deliberate in the procurement of capabilities — including data rights.

Project Linchpin picks up the torch from Project Maven and demonstrates both how to adapt AI technologies to military use and how the commercial sector can benefit from tackling hard defense problems first. The Department of Defense has already taken important steps to allow for agile acquisition, including the introduction of the Adaptive Acquisition Framework and the Software Acquisition Pathway that modernize the acquisition of software-centric products.

Project Linchpin will capitalize on the Software Acquisition Pathway while also informing new, AI-specific policies. And it will do so as a program of record, which is an important distinction. Research and prototyping can only go so far to meet the needs of the Army. Real value and return on investment in technological development only emerges when a capability is tested and validated for deployment onto a program available for operations. This transition from science and technology research projects to programs of record is often referred to as the valley of death because of the challenges projects face to meet fielding requirements. In the case of AI, new or retrained models must be delivered quickly during operations to counter adversarial attacks that can reduce performance. Combatants can poison data sets with inaccurate data, hide and camouflage objects to avoid detection, or even implant trojans to neutralize AI capabilities after they’ve been deployed. These threats put decision-making timelines in jeopardy and reduce the trust users have when employing them in operations. Project Linchpin seeks to provide a risk-management framework for AI that is incorporated across its entire acquisition approach — from supply chain to deployment — and ensures that AI is secure and reliable.

The Army’s Future: Pipelines

Machine learning operations pipelines can rapidly and continuously provide high-performing, trusted models. They are used by expert teams with training and experience in developing, evaluating, and deploying capability at a scale that comes from the centralized management of resources like data sets and high-performance computing. Project Linchpin, by managing machine learning operations for a compilation of programs like intelligence sensors, has brought this kind of pipeline to the Army, which will help in the future development of even more advanced AI capabilities. 

This system also enables the most innovative partners in the technology industry to create a dynamic marketplace for AI and machine learning capability development. Small, non-traditional contractors often struggle to break into the defense sector even if they have competitive or highly innovative technologies because they are unable to meet certain security or facility requirements. Sometimes, they lack the robust or exquisite data sets required to train and prove viability for operational use cases. Project Linchpin’s structure provides the development and testing environment necessary for the Army to test capabilities from vendors of all sizes. Combining the agile assessment of those technical tools with modular contracts will ensure that the Army can continue delivering more affordable and better-performing capabilities over time. 

The project will provide ample opportunity for collaboration with the technology industry given that there is the right balance of security, intellectual property protections, competition, and opportunity for growth. The Army can use this platform to drive government and industry standards for security and interoperability, which makes Project Linchpin a truly revolutionary tool for digital transformation. Adopting industry best practices for AI and machine learning pipelines creates dual-use opportunities and encourages an ecosystem of partners for the Army and the Department of Defense that can ensure that the United States maintains an advantage over adversaries in AI.  

It is not yet clear whether the current investment in machine learning operations tools and services will be significant and enduring enough to sustain this potential advantage. If the Army learns from the path that Project Maven paved, and if governance and interoperability keep pace with the technical and ethical challenges in the AI space, the Army will be able to keep ahead of its adversaries. Project Linchpin is the first program that aims deliver trusted and operationally relevant AI capabilities. However, as its name implies, it may serve an even more important purpose: ensuring that the Army develops strategic guidance about the responsible use of AI for the next generation of AI industry partners that can fuel the Army of 2030 and beyond. 

 

 

Maj. Nick Bono is an active Army acquisition officer, and the Department of the Army Systems Coordinator for Intelligence Systems, including Project Linchpin. He previously served as the assistant product manager for the Armys first AI-enabled intelligence ground station, the Tactical Intelligence Targeting Access Node (TITAN), and as a member of the Algorithmic Warfare Cross-Functional Team (Project Maven). He holds a Masters of Policy Management from Georgetown University and a Juris Doctor from the University of Tulsa College of Law.

Bharat Patel is the product lead for Project Linchpin. An engineer with the Armys Program Executive Office, Intelligence, Electronic Warfare, and Sensors, he managed the Armys pilot program with Project Maven to develop and evaluate AI for sensors. As the former chief technology officer for Project Manager, Intelligence Systems and Analytics, he coordinated science and technology efforts across programs and experiments like Project Convergence. He holds a Bachelor of Science in Computer Science from Rutgers University and a graduate certificate in Systems Engineering from Johns Hopkins University.

Image: U.S. Army photo by Sgt. Woodlyne Escarne

Related posts