Phantoms translating real-world operational constraints into innovative solutions Published April 13, 2026 By Capt. Amelia Leonard CAMBRIDGE, Mass. -- Innovation does not always begin with a new model or a new algorithm. Sometimes it begins when there’s a spark of ingenuity after experiencing or observing something that could seemingly be solved by a simple change in process or procedure. It can come from observing a scheduler spending hours building a daily flying schedule by hand, a cost analyst waiting weeks for fragmented contract context, or an engineer relying on low-fidelity design tools that can create expensive downstream revisions. Across the DAF–MIT AIA, Cohort 15 Phantoms are showing that one of the most important forms of innovation is translating real operational friction points and converting them into technical approaches that are usable, relevant, and mission-focused. Some of Cohort 15's projects start with familiar operational burdens and reframe them as solvable technical problems. Aaron Kenyon, Cohort 15 Phantom, decided to solve one of these very operational burdens. His MQ-9 Flying Schedule Optimization project addresses a manual training-unit workflow that can consume several hours per day for each scheduler while balancing instructor availability, aircraft and simulator inventory, syllabus progression, and training priorities. His project began with a deterministic baseline scheduler and slowly evolved into a reinforcement-learning approach supported by a graphical interface, translating a squadron-level burden into a practical decision-support tool designed to speed schedule generation and improve its outputs. “I compared what they built all day long versus what I generated in like five seconds, and it was virtually the same,” said Kenyon during his midpoint cohort presentation briefing held in March. The project’s stated bottom line is simple and operationally meaningful. It dramatically reduces the time required to build schedules. The same translation logic appears in projects where the operational problems are more technical but are nonetheless urgent. Ethan Jackman’s work on active learning in hypersonic multidisciplinary design optimization starts from a conceptual design process dominated by large volumes of low-fidelity data. That workflow can produce suboptimal designs that later require significant revision, which affects cost and scheduling. Jackman’s project translates that challenge into an integrated geometry-trajectory co-design framework that uses sparse high-fidelity computational fluid dynamics data, Gaussian process surrogate modeling, Bayesian optimization, and active learning to improve fidelity during conceptual design while keeping computational cost manageable. It is a strong example of translating a deeply specialized engineering problem into a more rigorous, decision-relevant process. His project takes an incredibly precise and technical problem and uses machine learning to allow the user to make decisions in a faster and more precise way. Other fellows are translating dynamic mission environments into adaptable AI architectures. Gavin McCormick’s modular expert-based classification project responds to a problem common across military and intelligence pipelines. His project addresses conventional classification systems which often require complete retraining when a new target class is introduced. This process often creates delays in fast-changing operational settings. His solution is to use a modular mixture-of-experts architecture with DINOv2 features and attention-driven fusion to support class-incremental updates without retraining the entire model. DINOv2 models produce high-performance visual features that are robust and perform well across domains without any requirement for fine-tuning, according to Github, an online cloud-based resource where members store, manage, share program coding. “We’re just trying to replicate our framework in the machine-learning context,” McCormick said. His project converts the reality of changing electronic warfare and radio frequency and signal intelligence environments into an AI program built for adaptability rather than static assumptions. Translation also matters in organizations where the bottleneck is not in the sensing or classification, but in bureaucracy and decision support. Paola Kefallinos’ AI-enhanced Independent Government Cost Estimate project addresses a process that can take months and is often constrained by fragmented historical data and the challenge of securing the right technical context. Her work translates that slow, manually intensive acquisition workflow into a predictive pipeline built from a decade of DAF contract data, narrowing hundreds of variables into a smaller, more usable feature set and applying machine learning methods to estimate potential award value. The aim is not just speed for its own sake, but faster and more rigorous cost insight that helps acquisition teams move out of a reactive posture. In addition, Cohort 15 Phantom, Marine Joshua Yin is working on a Multi-Agent Legal Reasoning System offering another version of the same idea but at the intersection of policy, law, and AI. Military legal advisors must navigate a dense and evolving body of statutes, regulations, and case law, where the difficulty is not just finding information but ensuring that it remains authoritative and properly cited. Yin’s project translates plain-language legal questions into structured legal memoranda backed by verified sources. “The obvious first instinct is to say, ‘Use AI.’ but there’s a domain problem,” Yin said during his midpoint briefing. “If you want to convince someone that you’re right, you have to quote exactly where the law says it,” he said. His project bridges that technical possibility and institutional requirement by using AI for synthesis while simultaneously preserving legal traceability and compliance. All of these projects showcase Phantoms distinctive value is not only technical experimentation. It is the development of Phantoms who can move between research environments and operational realities without losing sight of either. Whether the problem is scheduling, acquisition, hypersonic design, intelligence classification, or legal reasoning, the innovation lies in the translation itself. These Phantoms understand the mission well enough to frame the right problem, then shape their technical solutions to fit the real-world context in which it must be used. These projects are examples of what happens when operational experience, research methods, and mission needs are brought into deliberate alignment. These translations aren’t occurring in the abstract arena, but in the space between research and operations, where translation becomes the innovation.