Telefante Atlas
AI-assisted multimodal cultural archive infrastructure
Telefante Atlas
AI-assisted multimodal cultural archive infrastructure
Type: Clustering system / Knowledge infrastructure Start: 2026
The underlying problem
Cultural institutions hold archives they cannot truly explore.
The classical archive answers one question: Where is object X? It does not answer the question that matters most: What is beside it, that no one knew yet?
Search requires knowing what you are looking for. Classification requires predefined categories.
Neither produces genuine discovery.
Research question
Can a multimodal clustering system make visible the connections that no taxonomy has anticipated — conceptual genealogies, visual affinities, structural recurrences — within a heterogeneous cultural archive?
The conceptual reference is Aby Warburg's Atlas Mnemosyne: an archival principle that groups images not by category but by internal affinity — by the formal and emotional energy they share.
AI-Archive translates that principle into a parametric, digital infrastructure.
The dataset: Telefante Atlas
The first dataset is the studio's own archive: over 20 years of audiovisual and performance practice — material from Cologne, Amsterdam, Bogotá, Marseille.
Currently organised by metadata. Inaccessible in its real connections.
The Telefante Atlas is not only the technical test bed. It is the object of research: an archive that embodies exactly the problem the system is built to solve.
The process is part of the result.
How AI-Archive works
Multimodal embeddings — material is converted into meaning vectors: visual similarity via CLIP, semantic proximity via multilingual language models.
Unsupervised clustering — UMAP + HDBSCAN groups the material by characteristics extracted from the content itself, without prior categories, without external knowledge.
Exploratory interface — visual map of the archive, cluster zoom, relation lines, curatorial views. Curatorial authority remains human. The system extends the gaze — it does not replace it.
Process phases
Phase 1 — Preparation Digitisation and normalisation of the Telefante Atlas. Multilingual metadata. Vector database construction.
Phase 2 — Clustering and validation UMAP + HDBSCAN application. Iterative curatorial validation: do the clusters align with what a trained eye would see?
Phase 3 — Interface Development of the visual navigation layer. Stack: React / Svelte + D3.
Phase 4 — Pilot operation Test with an institution in Cologne or NRW. Documentation of results and adaptation.
What is at stake
An archive that today exists only as storage can become a living system of knowledge.
The effect is not only technical: it changes what questions an institution can ask about its own collection.