Enterprise Research Data Hub — Design Study

Read this first: This is a design study — a concept architecture I designed to show how I would approach unifying research and library data. It is not a built system, it does not describe any employer's actual systems, and it was not approved by any employer. It grows directly out of the real work in this portfolio: curating a research library collection, optimizing regulatory information systems, and hunting for information across disconnected sources on behalf of researchers.

The Problem It Solves

Research and library data accumulates in disconnected places: journal and article collections, regulatory documents and intelligence, internal reports, and the systems that hold each of them. Every source has its own structure, its own vocabulary, and its own way in. A researcher with one question searches several systems, reconciles several vocabularies, and still can't be sure they found everything — or cite cleanly what they found. The data exists; the architecture to treat it as one asset does not.

Architecture Layers

1 · SourcesLibrary collections, regulatory documents & intelligence, internal research materials

→

2 · IngestionScheduled + on-change intake, parsing, normalization into a common record shape

→

3 · Metadata & TaxonomyShared taxonomy, required metadata (source, owner, date, rights, version), deduplication

→

4 · Governed Store & CatalogOne catalog over the unified records; lineage back to every source preserved

→

5 · Governed SearchSingle search across sources, filtered by the searcher's own permissions

→

6 · AI AssistantAnswers questions from the catalog — every answer carrying citations to the records behind it

Identity & Access (cross-cutting)Enterprise SSO throughout; the hub never widens access — a user finds only what the source systems already allow them to see.

Governance & Audit (cross-cutting)Ownership, retention, and versioning rules per source; a log of what was ingested, searched, and cited for review.

The Governance Thinking

The metadata layer is the product. Unification fails without shared vocabulary: one taxonomy, enforced required fields, and stewardship for terms — otherwise the hub is just a bigger pile.
Every record keeps its source. Lineage is non-negotiable in research and regulatory work: any result must trace back to the exact source record and version it came from.
Access follows the source, not the hub. Permissions are inherited and enforced at search time; the hub is a window, never a back door.
Versions and retention are designed, not patched. Regulatory material changes; the architecture treats superseded versions as first-class — retained, marked, and excluded from default results.
Quality is measured at ingestion. Missing metadata, duplicates, and stale records are caught and reported when data enters, not discovered when a researcher fails to find something.

The AI Layer — Citations or Nothing

The assistant layer answers only from records in the governed catalog, and every answer carries citations to those records. If the catalog doesn't support an answer, the assistant says so and points to the closest material instead of filling gaps from general knowledge. In a research and regulatory setting, an assistant that can show its sources is useful; one that can't is a liability — so grounding is the architecture, not a feature bolted on.

How Success Would Be Measured

Findability

Time-to-find for a known item before vs. after; percentage of searches ending in a result the researcher opens — the hub wins only if one search replaces several.

Metadata coverage

Share of records meeting the required-metadata standard, tracked per source — the health metric for the taxonomy layer.

Cited-answer quality

Share of AI answers whose citations fully support the claim, sampled and reviewed by librarians — trust, measured instead of assumed.

Adoption

Active researchers using the hub as their first stop, and repeat usage over time — voluntary adoption is the honest signal, the same one I trust in learning programs.

These are proposed measures for a concept — this study reports no results, because the system has not been built.