2026-11-03 –, Beavis
Large Language Models are changing geospatial science, but AI agent development starts with ad hoc prompting instead of structured design, producing agents hard to reproduce or evaluate. We share a specification-first co-design methodology for community-maintained geospatial agents.
Large Language Models are rapidly changing how scientists interact with geospatial data. While many demonstrations showcase impressive AI capabilities, building reliable scientific AI agents remains difficult because development often begins with prompt engineering instead of structured design. The result is scientific agents that are difficult to reproduce, extend, or evaluate and a fragmented landscape where useful work rarely outlives the demo that produced it.
This talk presents a collaborative and open methodology for co-designing trustworthy AI agents for geospatial science, and just as importantly makes the case for treating agent specifications, prompts, and benchmarks as shared community infrastructure rather than one-off project artifacts. Rather than focusing on prompt engineering alone, the methodology uses a specification-first workflow that captures scientific requirements before implementation, built on collaborative design between subject matter experts, developers, and AI helper agents. This produces reusable engineering artifacts scope and stakeholder definitions, workflow decompositions, tool specifications, reasoning policies and guardrails, and evaluation plans that improve transparency, maintainability, and reproducibility across the community.
We will walk through this methodology using real geospatial use cases developed with the AKD-Labs Agent Co-Design Studio, including example of Earth-observation data-search agent illustrating how the same design pattern generalizes across problem domains.
The talk will also focus on the community-building dimension of this work: our effort to publish specifications, prompts, benchmark templates, and documentation to an open AI-for-Science GitHub repository. We'll discuss how this repository is meant to function as shared, extensible infrastructure not a static archive and invite discussion with attendees about governance, contribution models, and what it would take to grow this into a durable open-source ecosystem for trustworthy geospatial AI, similar in spirit to other community-maintained standards and toolchains in the FOSS4G world.
Dr. Nidhi Jha is a Research Scientist at the University of Alabama in Huntsville, supporting NASA's Office of Data Science and Informatics (ODSI). Her research focuses on geospatial artificial intelligence, remote sensing, Earth observation, and AI-driven scientific discovery. She is the SME lead for the Accelerated Knowledge Discovery (AKD) project initiative, where she develops collaborative methodologies for designing trustworthy AI agents for science. Her work spans agentic AI, Earth science data discovery and the integration of large language models into scientific workflows. She is actively involved in advancing open, community-driven frameworks for AI agents in Earth science and collaborates with NASA and the broader research community to accelerate scientific discovery through trustworthy and reproducible AI.