Vibe-coding ROMAD-AI: Cadet builds military app with chatbots

A U.S. Air Force cadet built a prototype application called ROMAD-AI using "vibe-coding," a process that relies entirely on prompts to generative AI chatbots, over a three-month period. Joshua Lynch used paid versions of Anthropic's Claude, OpenAI's ChatGPT, and Google's Gemini, and finished a document-processing prototype via Google AI Studio App.

How did a novice build ROMAD-AI with chatbots?

Lynch began with no prior coding skills and relied on chat-based prompts to generate and refine code, working primarily through the chatbots' main web chat functions rather than inside an integrated development environment. He split problems into small parts, framed questions clearly, and steered conversations back on topic when the assistants strayed; those tactics helped mitigate common failures such as lack of hierarchical focus and unintended edits to unrelated code.

Lynch and his mentor Laura Niss observed the process while the project ran under the U.S. Department of the Air Force–MIT AI Accelerator Phantom Program. Most development used Claude, ChatGPT, and Gemini; the final application was produced using Google AI Studio App, which can create applications that interface with the Gemini API and has AI integrated in the development environment.

What did ROMAD-AI do and how was it scoped?

ROMAD-AI started as a plan for a tactical tool offering target recognition, modular intelligence, surveillance and reconnaissance, autonomous striking, and battlefield communications management, but limitations in the AI and available development time led Lynch to re-scope the project. The delivered prototype performs basic document processing, analyzing tactical maps and generating mission-planning documents through an interface with a vision-language-model-powered chatbot, rather than providing the full autonomous battlefield functions originally envisioned.

The team found that recognizing chatbot limitations and learning to work around them consumed most of the timeline. Niss says the final prototype demonstrated usefulness for prototyping and communication between nontechnical experts and engineers: "I was quite impressed with this final product, and it showed me how powerful these systems can be at prototyping designs from nonexperts." The project also surfaced operational risks: Lynch did not realize the final application was sending input documents to a Gemini model for analysis instead of parsing them locally, a behavior that raises security concerns for sensitive workflows.

How did the chatbots compare in use?

Lynch and Niss tracked perceptions of different systems over time and across system updates, finding that Claude showed more stability than ChatGPT on traits such as likeability, anthropomorphism, and perceived intelligence. Lynch reported that the chatbots served as helpful tutors but sometimes produced inaccuracies on topics he knew well, reinforcing the need for expert review and code vetting.

The research concluded that while AI chatbots can empower nontechnical service members to produce viable software prototypes for mission problems, they are better suited as prototyping assistants than as production tools for sensitive or critical applications. The team noted code review and security vetting remain bottlenecks when AI generates functional code.

Why it matters

The project shows that nontechnical personnel familiar with a problem space can use generative chatbots to turn ideas into working prototypes within a short time frame, here three months. That capability shortens the path from problem identification to a testable design and can improve communication of requirements to technical teams. At the same time, the example where the prototype sent documents to an external model underscores an operational risk: prototypes that rely on cloud model calls may leak data unless development enforces secure, local processing and rigorous code review.

What to watch

Monitor whether teams move prototypes away from cloud model calls toward local parsing or vetted APIs, and watch for adoption of development workflows that require explicit code review and data-flow checks. Also track outputs from the Department of the Air Force Artificial Intelligence Accelerator, which sponsored the research under Cooperative Agreement Number FA8750-19-2-1000.