Maintaining Scientific Rigor in AI-Assisted Development: A Validation-Focused Methodology
The rapid evolution of AI-assisted development—spanning concepts like "vibe coding," the Model Context Protocol (MCP), and autonomous agents—presents a paradigm shift toward "Agentic Engineering." While these tools offer immense productivity, their application in scientific computing introduces unique failure modes distinct from standard software engineering, particularly regarding numerical precision, subtle physics violations, and reproducibility. General-purpose AI agents are powerful, but they require domain-specific adaptations to ensure reliability in high-stakes scientific workflows.
BSSw Fellow Victor Weeks will establish a comprehensive methodology for "Scientific Agentic Engineering," defining how to integrate these technologies into rigorous workflows without compromising scientific standards. Key deliverables include a tool-agnostic framework for responsible AI adoption and a Reference Workflow—a structured exercise in which participants use AI tools to build a simple weather model error statistics application, applying validated scientific agents, skills, and templates designed to prevent the failure modes that general-purpose tools routinely produce. Through an interactive tutorial series and community workshops, the project will equip the scientific software community with the shared vocabulary, practical methodology, and domain-specific tooling necessary to transform broad AI capabilities into reliable scientific software development instruments.
Victor is a Software Engineer at the NSF National Center for Atmospheric Research (NCAR), specializing in robust workflows and infrastructure for numerical weather prediction systems. Working across laboratory and program boundaries, he supports diverse scientific projects ranging from atmospheric chemistry to ML-based prediction. Victor is an early advocate for Agentic Engineering within his lab, actively integrating AI coding agents to enhance development velocity without sacrificing quality. Previously, he worked at STAR, LLC, maintaining mission-critical WRF systems for international government and defense clients. An alumnus of Colorado State University (Computer Science, Psychology), Victor is dedicated to building software that is as rigorous as the science it supports.
