RDA 27th Plenary Meeting, London, 6–8 October 2026. Poster 3, Tuesday 6 October, 18:00–19:30, Kia Oval.
Poster presentation and companion materials
A persona-based agentic AI blueprint and application for FAIR research data, metadata, and schema and crosswalk workflows
Six persona-based AI agents, each modeled on an established research-data professional, transform a researcher’s raw dataset into a FAIR-compliant, six-file deposit package, with human judgment and transparency governing consequential decisions.
Organizing agentic AI by professional persona brings deeper intelligence to research-data workflows, making them FAIR by design, legible by construction, and accountable at irreversible steps.
Research data curation is specialized, labor-intensive work. A FAIR-compliant deposit requires a schema-conformant metadata record, persistent identifiers, licensing and documentation, an auditable provenance record, and crosswalks to adjacent standards.
Few researchers have the time or specialized training to meet these requirements, and few institutions can staff every requisite role. As a result, many datasets are deposited without being findable, interoperable, or reusable, and fall short of funder data-management and sharing requirements.
The Research Data Director addresses this gap: a coordinated team of persona-based AI agents produces a vetted first draft of the complete deposit documentation for human review.
An agentic system pursues a goal: it plans, uses tools, checks its own work, and repeats until the task is done. There are two principal ways to specify such a system, and the second builds on the first.
Starts from what the system must do.
Discrete functions (search, extract, validate, crosswalk, export), each with inputs, outputs, and checks, arranged in a layered reference architecture. Auditable, testable, and straightforward to govern; its characteristic risk is brittleness at the edges.
Example: the RDA/Microsoft Data Director Agentic AI Tool Blueprint, endorsed by RDA in September 2026.
The persona model builds on the capabilities model
Starts from who would do the work well.
In the inner loop, the Metadata Librarian generates; the Fact Checker and FAIR Expert critique in parallel; the Director decides, with at most two rounds before human escalation.
Each capability is paired with a role, expertise, and judgment, and assigned to a synthetic professional. The result is a team that can deliberate and execute.
Personas are context-rich and legible to domain specialists. Their characteristic risk is being persuasive but wrong, which is why every persona carries a defined human checkpoint.
Real agentic AI systems synthesize both: synthetic persona-based systems execute with judgment and expertise.
Each agent is modeled on a professional research-data role and specified by its professional analogue, scope of competence and bounded authority, principal output, escalation trigger for human review, and multi-agent pattern.
| Agent | Professional analogue | Principal output | Escalates to human review | Pattern |
|---|---|---|---|---|
| Research Data Director | Head of Data Services | Dispatch and project-management plan; approval gates | Irreversible actions, such as DOI minting | Orchestrator |
| Research Data Librarian | First-contact intake specialist; sensemaking | Working Brief (JSON) | Sensitivity and privacy flags | Sequential intake |
| Metadata Librarian | Senior cataloger; taxonomy and ontology | DataCite, EML, or DDI record; crosswalks and mappings | Low-confidence schema mappings | Generator |
| Fact Checker | Research integrity specialist | Verification report | Unresolvable external sources or APIs | Evaluator (parallel) |
| FAIR Expert | FAIR assessor | F-UJI metric score | Sub-threshold FAIR scores | Evaluator (parallel) |
| Provenance Recorder | Digital archivist | PROV-O audit log of every action | None; continuous ambient audit | Continuous audit |
Four secondary agents join the team when a case requires them.
The Persona Appendix (Appendix F) specifies each agent with a behavioral layer and a psychoanalytic sense-making component, to better answer questions and serve user needs.
Agents execute, read the shared document (memory) freely, and propose logical next steps, challenges, and possibilities.
Serious questions and irreversible acts, such as minting an identifier or publishing to a permanent record, belong to a human evaluator.
The boundary principle, adapted from the Research Data Director blueprint
Human oversight is a property each agent carries individually, not a global setting. The following triggers pause the system and route the decision to a named human role. The list is representative; institutions should extend it for local policy.
| What stops the system | Raised by | Routed to |
|---|---|---|
| DOI minting, license assignment, or public deposit (always gated) | Research Data Director | Researcher and research data librarian |
| CARE flag: Indigenous, community, or sovereign data signals detected | FAIR Expert | Community representative; research ethics office |
| Personal identifiers or sensitive location data found in files | Research Data Librarian | Researcher; data protection officer |
| Crosswalk field marked low-confidence or unmappable | Metadata Librarian | Disciplinary metadata specialist |
| Identifier claim cannot be verified by any API | Fact Checker | Researcher; research integrity officer |
| FAIR metric below institutional threshold after two revision rounds | Research Data Director | Research data services lead |
| Final deposit diverges from stated intake intent | Research Data Director | The researcher |
The Thinking Window exposes agent deliberation as it happens, so outputs can be reviewed as one would review a professional colleague’s work. It is opt-in and fully archivable, and it lets the researcher intervene directly in the workflow.
A single raw dataset yields six deposit-ready outputs. The researcher submits once; the repository receives everything it needs.
| Document | Contents | Immediate value |
|---|---|---|
| Metadata record | Complete record in the target schema (DataCite 4.6 plus domain schema), with required and recommended fields populated, sourced, and confidence-rated | Findable in repository search and Google Dataset Search |
| Crosswalk table | Field-by-field mapping to at least two adjacent schemas, with confidence ratings and human-review flags | Findable from adjacent disciplines; enables synthesis |
| README file | Purpose, methods, variable descriptions, known limitations, loading instructions | Reuse without prior knowledge of the dataset |
| Data dictionary | Variable-by-variable documentation: name, type, units, controlled vocabulary term and URI, missingness codes | Satisfies FAIR R1 and I2; enables machine parsing |
| FAIR assessment | Score per F-UJI metric, with a plain-language explanation of each failure and remediation steps | Demonstrates FAIR compliance to funders |
| Provenance sidecar | PROV-O log of every agent action, tool call, model version, and human decision | Reproducible, auditable publication process; FAIR R1.2 |
The Metadata Librarian maps each native record, field by field, to adjacent schemas and assesses confidence for every mapping. Low-confidence mappings are escalated for human adjudication, and the crosswalk is exported as a first-class artifact for discovery. Treating crosswalks as governed, human-reviewed outputs addresses the FAIR principle most resistant to reliable automation: interoperability.
The figure below makes the argument through an analogy from precision agriculture, where several instruments observe the same orchard. Each panel is shown separately beneath it.
A plant physiologist cannot diagnose a failing irrigation emitter by eye. Four independent instruments each record one property of the same tree: near-infrared reflectance flags canopy stress; the thermal camera shows the canopy running 2.3 °C warmer; the soil probe reads 14% moisture at 25 cm against 22% expected; the GPS grid fixes the row and tree.
Only when the four are aligned on the same coordinates and time does the diagnosis appear. Metadata works the same way: each schema describes one facet of a dataset, and alignment is what makes the whole findable and interpretable.
Every instrument reveals something and is blind to something else. The RGB camera captures geometry but not biochemistry; the thermal camera sees heat anomalies but cannot distinguish their causes; LiDAR resolves structure beneath the canopy but not chemistry; ground sensors give continuous truth at a point but cannot extrapolate across space.
Metadata schemas behave the same way. DataCite, EML, and DDI each capture what their discipline needs and omit the rest. A crosswalk records exactly where their fields meet, and where they do not.
When distinct instruments are aligned in a shared spatial and temporal framework, the relationships among their measurements encode information that resides in no single measurement.
Crosswalks are that shared framework for metadata. They let data from one discipline be found, read, and combined by another, forming larger bridges and analogues between schemas and enriching the possibilities for scientific discovery.
Foundational blueprint
The architecture, personas, outputs, and human checkpoints. Research Data Alliance, 2026.
Technical companion
Agent specifications, shared memory, human checkpoints, and open protocols for implementers.
Persona construction
The persona team in detail, re-engineered for current persona practice, with working profiles for each agent that can be customized to a domain or institution.
Capabilities model
The capabilities-based blueprint of the RDA Global Agentic AI Initiative, sponsored by RDA and Microsoft, on which this work builds. Endorsed by RDA, September 2026. The author was a contributing working-group member.
A separate project that applies the architecture to multimodal biomedical data, with its own early-stage prototype.
Domain variant
RDA VANTAGE Working Group, August 2026, with its background paper Diagnostic Synthesis: Biomedical Research Data, Metadata, and Discovery Through Imaging, Genomics, Sensors, and AI.
Introduction to agentic AI
An introduction to agentic AI and to capabilities-based and persona-based models, written for newcomers. UC Riverside Libraries, Research and Technology Services, AI Literacy Series, September 2026.
Background to the crosswalk examples
The convergence of multi-scalar remote sensing and computational analysis in scientific inquiry, and the origin of the instrument analogy used for crosswalks.
Governance companion
A framework for AI alignment, safety, ethics, and transparency in human–AI research: tiered human checkpoints, verification reports, and validation of DOIs, author identifiers, and provenance. RDA Agentic AI Working Group, Research Data Alliance, September 2026. Its Appendix C, the Verification Kit, applies three-layer, cross-model verification with human adjudication to a worked example, Unlocking the Invisible Mile.
At a glance
One page that puts the governance layer into practice:
The blueprint pairs a conceptual architecture with a buildable implementation, realized through twelve prototype iterations in July 2026. All versions open in the browser.
Agents reason live over the researcher’s description and dataset summary.
The six-persona team with transparent deliberation. Version 11 adds a demo mode that runs in about 90 seconds with no input needed.
An experiment in spoken interaction with the persona team; requires an ElevenLabs connection.
See the annotated prototype gallery: thirteen screenshots with notes
Early-stage prototype, August 2026
A separate project applying the architecture to multimodal biomedical research data: a metadata and crosswalk studio for dataset registration, metadata normalization, provenance and quality assessment, and human-reviewed crosswalks across clinical and imaging standards. Synthetic or de-identified data only; not for clinical use.
RDA Artificial Intelligence and Data Visitation (AIDV) Working Group, 13 July 2026.
From simple office productivity to research-level agentic teams. UC Riverside Orbach Science Library AI Literacy Series, 24 September 2026.
Research Data Alliance US Cohort Participation, California Digital Library Briefing.
Each overview approaches the project from a different angle and introduces a specific paper. They were generated with Notebook LM from the source papers and are designed as accessible introductions for undergraduates, advanced high school students, and colleagues new to the subject. A good place to start: Diagnosing the invisible, then the three-minute overview of metadata and corporations.
Introduces Spatial Synthesis (PDF), the source of the crosswalk analogy in section 7, and the economic case for the Research Data Director.
Video, 9:41 How aligning GIS, satellite, drone, and ground data reveals what no single instrument can see.
Innovation, economics, and opportunities
Video, 2:40 The economic side of metadata and data synthesis, and the opportunities agentic AI opens for organizations.
Introduces the Research Data Director blueprint through a worked case.
Video, 8:13 How an epidemiological dataset meets the synthesis bottleneck, and how the Research Data Director’s crosswalks and review steps address it.
A variant application. Introduces the biomedical blueprint and its background paper, Diagnostic Synthesis.
Video, 7:39 Why medical data sources describe the same patient in different vocabularies, and how governed crosswalks make discovery across them possible.
Introduces the RDA/Microsoft capabilities blueprint, developed in the RDA Global Agentic AI Initiative, April to June 2026.
Video A simplified overview of the capabilities model.
Video An advanced overview of the capabilities model.
Two longer, conversational audio discussions introducing ideas regarding scientific research and possibilities with the Research Data Director. They suit listening on a commute or a walk. Choose a playback speed, or download a file to listen on your phone or computer in any audio app.
Audio, 30:58 Why research data stays siloed, and how agentic AI can connect it.
Audio, 20:04 How agents translate data between disciplines, schemas, and communities.
This work originated in, and builds on, the RDA Global Agentic AI Initiative Data Director cohort, sponsored by RDA and Microsoft. Between April and June 2026, the initiative convened some 90 participants worldwide in an eight-week programme, consolidating into two regional cohorts of about 16 data librarians, research scientists, metadata specialists, and agentic AI developers, whose community brought together the domain expertise and multidisciplinary perspectives underlying research-data lifecycle workflows. The author continues to participate in the RDA capabilities implementation group.
Many of the ideas also took shape through the University of California, Riverside, Orbach Science Library GIS and AI Literacy Series, a set of workshops on AI, GIS, data, and drones organized by Janet Reyes, and through the RDA Agentic AI US Cohort meeting in Washington, which sparked much of their coalescence.
With gratitude to Connie Clare (Community Coordinator, Research Data Alliance, UK); Becky Grady and Maria Praetzellis (California Digital Library, US); and AIDV and VANTAGE colleagues Patricia Buendia, Seonyoung Kim, and Francis Crawley for their support of this project.
Uzwyshyn, R. J. (2026, October 6). The Research Data Director: A persona-based agentic AI blueprint and application for FAIR research data, metadata, and schema and crosswalk workflows [Poster presentation]. Research Data Alliance 27th Plenary Meeting, London, United Kingdom. https://rayuzwyshyn.net/rda-p27/