Digital Research Environment

The Digital Research Environment (DRE) is the digital infrastructure unit of the Africa Multiple Cluster of Excellence at the University of Bayreuth. We design, build, and maintain the data systems that connect researchers across the Africa Multiple Research Centres (AMRCs) and partner institutions worldwide. The work spans research data management, knowledge-graph development, AI-assisted data processing, and digital literacy, all grounded in the FAIR and CARE principles and a commitment to multiple knowledge systems. I joined as Data Curator in 2026.

Two principles guide how we work. Curation and description are a joint effort with partners at the AMRCs in Ouagadougou, Makhanda, Lagos, and Eldoret: each centre describes the data it knows best, and the DRE runs the shared infrastructure that connects them. Storage, in turn, stays distributed by default. Data remains in its local repository, while Bayreuth holds the metadata layer that points to it, so research data becomes findable without being relocated. Both commitments are design constraints as much as principles, and much of what follows is an account of what they cost and what they make possible.

All of this metadata comes together in the Africa Multiple Interactive Research Atlas (AMIRA), an Omeka S site that makes the collections of the AMRCs and their partners searchable from one place (Madore, 2026) . It is the visible layer of the team's work, and the testbed for most of the questions below: nearly 4,000 research items, more than 90 projects, some 1,600 people and 600 organisations, in 28 languages across 42 countries.

Describing data across five centres

Federated description is the premise the rest depends on. The AMRCs and partner institutions hold substantial research collections and datasets; the DRE works with each centre to assess the scope, nature, quality, and size of that data, and to determine what can be linked to the platform. In practice this means structuring, describing, and quality-checking collections so they stay findable and reusable; running consultations with projects from the start, so data management plans are drafted while decisions are still cheap to change; and preparing datasets for deposit with DOIs and long-term archival.

The research question underneath the service is whose description counts. High-quality metadata makes a researcher's data more discoverable and their work more visible to a wider community of peers, but the vocabularies that carry it are rarely neutral. Subject description most of all: where terms exist at all, they tend to derive from Library of Congress Subject Headings, a vocabulary poorly suited to African sources and knowledge systems. Building taxonomies in African languages, and mapping project vocabularies onto them rather than onto an inherited standard, is where the Cluster's commitment to multiple knowledge systems becomes a concrete modelling problem instead of a statement of intent.

Modelling African knowledge

Tagging entities is where most collections stop. The harder step is to model the relations between them, so that a set of records becomes a graph one can reason over: people linked to projects, projects to places, places to the subjects that recur across them. Reconciliation with Wikidata anchors those entities outward, but its African coverage is thin, which turns the work into a contribution rather than a lookup. Where no record exists, the entity has to be created.

Two problems make this more than a data-entry exercise. The first is accuracy: generic models systematically misfire on African persons and places, mistaking variant spellings, honorifics, and non-European name ordering, and an entity-linking pipeline is only as good as its disambiguation on exactly the cases the training data underrepresents. The second is coverage. Alongside the research-item metadata, the Cluster's own scholarship is a corpus no one has yet read computationally: extracting entities, places, and themes from the full texts, and modelling the relations among them, produces a map of what Africa Multiple's research actually covers, and what it overlooks.

AI-assisted processing

Researchers return from fieldwork with hundreds or thousands of scans, photographs, and recordings, and the material often stays unprocessed for years. AI-assisted processing is what makes that backlog tractable: OCR for printed documents and handwriting, transcription of audio and video, summarisation of large corpora, named-entity extraction, and automatic tagging and classification of digital collections.

Two constraints shape how we build it. The models run locally, on open-source weights and the Cluster's own hardware, which keeps the pipeline GDPR-compliant for material that is frequently special-category personal data, and keeps it affordable for centres that cannot absorb commercial API costs. And every machine-generated value carries its provenance: the model that produced it, when, and the fact that a machine produced it at all. Where traditional OCR announces failure through obvious garbling, LLM-based processing hides fabrication and silent normalisation behind clean text. Recorded provenance and human validation are what keep that trade honest.

Access, visualisation, and dissemination

A collection that cannot be interrogated is not findable in any sense that matters. Alongside searchable, public-facing collection databases and project websites, the DRE builds interactive dashboards, network graphs, and maps that make research data explorable rather than merely retrievable. AMIRA's own visualisations follow the atlas's premise: rather than discrete records, they foreground connections, clustering the subjects that co-occur across the collections and drawing the collaboration networks that no single catalogue record makes visible.

How scholars arrive is changing too. A growing share now ask an AI assistant a question in their own language rather than searching a catalogue with keywords and facets, while automated crawlers harvest open collections in volumes heavy enough to take the serving infrastructure offline. The institution that assembled and described the data is then neither cited nor represented well in what the models say about its subject. Between closing a collection off and leaving it to unrestricted extraction lies a mediated layer, in which the institution defines how a machine may query its holdings, the data stays on its own servers, and every answer carries a link back to its source. Building and evaluating that layer for AMIRA is the current strand of this work.