PhD researcher at the Computer Vision Center, UAB. I build search engines for historical archives that are comprehensive, measurable and relational.
Socrates asked Protagoras to cut up his answers into shorter pieces, to strip the rhetoric off his speech. Two and a half millennia later, we have not moved beyond that tension between discourse, truth and measure.
Search engines are increasingly being set aside, replaced by the rhetoric of large language models.
They privilege what sounds persuasive over what is true. Retrieval offers evidence, without the relationships between pieces of evidence.
Conversational interfaces suppress the sources a user will never consult, and with them, the user's own doubt.
A historical truth is only as precise as the sources retrieved.
Incorporate as many relevant sources as possible, to construct the most complete description of a historical phenomenon.
Anchor every source to identifiable, verifiable fragments of information, so the contribution of each one can be tracked.
Natively support inter-relationality: let people explore grounded connections between sources, not just primary sources alone.
It is tempting to think nobody has trained one yet, that it is a matter of time and money. It is not: historical documentation, as a whole, inherently resists large-scale pattern recognition pre-training. What it needs is a proxy, a system that behaves foundationally across domains, languages and typologies while remaining narrow in scope.
Born in a less communicated, less globalized world, historical documents exacerbate the differences between collections of similar domains.
Collections are sparse and incomplete; expert knowledge is required even for foundational tasks such as reading text.
Heritage carries cultural significance. Centralizing it outside each archive's data governance is legally and ethically out of reach.
Even when augmented with retrieved content, chatbots cannot count relationships, constrain literals such as dates or names, or account for the relevant content they never retrieved. I index photographs and documents through visual routines, graphs and ontologies instead.
Building one looks like a chimera, yet practitioners still need an initial seed for zero-shot prediction or light fine-tuning. I study decentralized learning and model editing to bypass centralization while improving low-resource, out-of-domain reading.
A date alone cannot index a specific fact, but temporality is entangled with the topics and styles of a period. It can serve as a proxy to relevant information from the preservation perspective.
Should a late nineteenth-century car and a modern one share a representation? Time evolves in a continuum, so models should read historical data as a progressive evolution of features, not as isolated, well-defined clusters.
No record matches. A chatbot would have answered anyway.
Indexing historical photographs and documents with factual data, within the project Graph Reasoning for Automatic Interpretation and Learning. In collaboration with the Xarxa d'Arxius Comarcals and the Centre d'Estudis Demogràfics.
Applied research and proofs of concept in image analysis for archives, under the agreement with the Xarxa d'Arxius Comarcals.
Metadata de-duplication for the music industry.
Science outreach: a (now defunct) Twitter bot that dated users' photographs.
Retrieval through continuous approximations of discrete ranking functions.