The virginia polytechnic institute and state university libraries will collaborate with the virginia tech sanghani center for AI and data analytics, and northwestern university libraries to research…

The Institute of Museum and Library Services awarded Virginia Polytechnic Institute & State University $1.226 million on September 1, 2026, under the National Leadership Grants program (CFDA 45.312) to research privacy-preserving artificial intelligence architectures for digitized library collections.

Virginia Tech Libraries, in collaboration with the Virginia Tech Sanghani Center for AI and Data Analytics and Northwestern University Libraries, will investigate how agentic AI systems combining structured reasoning with tool-based verification can address inconsistent descriptive metadata across digitized collections while mitigating privacy risks and institutional control concerns inherent in proprietary AI models. The project will produce a validated design framework for privacy-preserving AI, a benchmarking methodology for evidence quality, and an open-source toolkit for nationwide institutional adaptation. The work will generate a documented reference architecture that separates curatorial policy from model choice, enabling institutions to adopt improved AI models without redesigning metadata workflows. The project will also develop an open library of agentic skills—including prompt templates, workflow sequences, and grounding configurations—designed to process complex archival materials and reduce factual errors in metadata creation.

Performance of work is located in Blacksburg, Virginia, with a period of performance through August 31, 2029.

The virginia polytechnic institute and state university libraries will collaborate with the virginia tech sanghani center for AI and data analytics, and northwestern university libraries to research the challenge of inconsistent descriptive metadata across digitized collections and the risks to privacy and institutional control inherent in proprietary artificial intelligence models. This research will investigate how agentic artificial intelligence architectures combining structured reasoning with tool-based verification can be designed to meet library standards for transparency and local stewardship, producing a validated design framework for privacy-preserving artificial intelligence, a benchmarking methodology for evidence quality, and an open-source toolkit for nationwide institutional adaptation. The project will produce a documented reference architecture that separates curatorial policy from model choice, allowing institutions to adopt improved artificial models over time without redesigning their metadata workflows. The project will develop an open library of agentic skills (prompt templates, workflow sequences, and grounding configurations) designed to process complex archival materials and reduce factual errors in metadata creation.