The first generation digitized the corpus · The second faces judgment

Digital architect
for humanities

I design AI infrastructures so that complex knowledge can be queried, verified, and traced — without algorithmic fluency destroying the difference between evidence, interpretation, and absence of basis.

PhD in Communication · UPF Barcelona Degree in Philosophy · UAM Madrid NLP Developer
Scroll

Language models fabricate citations with the same fluency with which they get things right.

In knowledge domains — philosophy, psychoanalysis, textual research — this is not just a precision error. It is the destruction of the difference between what can be cited, what can be oriented, and what should lead to abstention.

Generic AI assistants respond with authority even when evidence is weak, ambiguous, or entirely nonexistent. Ateneo addresses this problem with an architecture that treats abstention as a core system capacity, not as a failure.

78–90%
Citations fabricated by GPT-4o
<40%
Accuracy on simple factual questions
86%
Bibliographic references invented
Sources: Allen Institute for AI / OpenScholar (2024) · OpenAI SimpleQA · University of Queensland
Main project

Ateneo — AI-assisted editorial research instrument

Not another chatbot with a corpus. It is an architecture where citation verification is deterministic, documentary location is auditable, and the system explicitly distinguishes between what it can support and what it cannot.

Citation Mode

Canonical anchor

Pure corpus search, zero LLM. What it returns cannot be fabricated — it comes directly from the database with page and document. Shielded by 7 verification guards.

Source Mode

Documentary location

It does not interpret. It helps find where a formulation is documentarily supported. Tracing back to corpus and source, with auditable traceability.

Abstention

Knowing when not to answer

The core capacity. When there is not enough basis, the system does not fill the gap with fluency. It marks that there is no evidence. This is what almost nobody does today.

Portability

Stoic as proof of method

Marcus Aurelius, Epictetus, Seneca. Not a second product: the demonstration that the same pattern of contract, evaluation, and validation works in another domain.

80/80
Production validations
(static + live)
4
Formal contracts
with 20-case test batteries
7
Verification guards
before validating a citation
2
Domains operating
Lacan + Stoicism
Try Ateneo → Framework Review
Architecture

Two floors, one criterion

Los modos que acabás de ver — Cita, Fuente, Abstention — forman el suelo firme. Encima, Ateneo permite una lectura abierta pero anclada. La apuesta es mantener apertura interpretativa sin perder anclaje textual.

Upper floor · Assisted reading

Anchored interpretation

Open reading that does not pretend to replace the researcher. It proposes articulations supported by visible evidence, with explicit limits when support falls short.

— MAP Mode · conceptual axes — Visible evidence + explicit gap — Reference roles (analyzer, contextual)
Solid ground · Verification

Auditable citability

Deterministic citation verification and documentary location. What comes out of this layer cannot be fabricated — it comes directly from the database.

— CITATION Mode · pure search + 7 guards — SOURCE Mode · documentary location — Abstention cuando no hay soporte
First week after launch

What early users said

Institutional
«Supera, o eso me parece, los productos de los que hablaba a Pastora como el Thesaurus Linguae Latinae.»
Ángel Delgado
Director de Soporte al Aprendizaje e Investigación · Biblioteca/CRAI · Universidad Pablo de Olavide, Sevilla
Institutional
«Me parece magnífico tu Ateneo, ¡pero magnífico! ¡Nos va a ahorrar tanto trabajo y tanto tiempo!»
Dra. Cecilia Gorodischer
Docente Investigadora, Posdoctora · Universidad Nacional de Rosario
Academic
«Una exactitud tajante, una lucidez instantánea y despiadada.»
Dr. Alberto Ruiz de Samaniego
Profesor Titular de Estética y Teoría de las Artes · Universidad de Vigo · Comisario del Pabellón Español, 52ª Bienal de Venecia
Public
«Esto es una LO CU RA. Sigan a Pablo que está diseñando una app que va a dar que hablar y mucho.»
Flavia Pittella
Escritora y creadora de contenido · 300K+ seguidores
Pablo Martínez Samper
Trajectory

Humanities + code,
without giving up either

I come from philosophy and humanities research — and I arrived at natural language processing because I needed to solve a real problem: verifying citations against original texts in a corpus where LLM hallucination is unacceptable.

That forced me to build text processing pipelines, hybrid search systems (embeddings + FTS + trigram), parallel corpus management in 4 languages, and retrieval evaluation with real metrics. 125 SQL migrations. A production system that works.

That journey is what makes Ateneo hard to replicate. In STEM, verification has crutches: DOIs, structured APIs, rich metadata. In humanities, the corpus is ambiguous, editions vary, attributions circulate without a primary source. If the method works on Lacan — 80/80 — it has good reasons to work on any complex corpus.

2003
Philosophy
UAM, Madrid
2020
PhD in Communication
UPF, Barcelona
Current
Lecturer · Dissertation Supervisor
UNLP (Argentina) · UNIR (España)
2024–hoy
Ateneo · NLP + IA
Production
Book · 2024
Ascenso y caída del padre
A history of the father figure through cinema. Xoroi Edicions. ISBN: 978-84-127324-8-1
Documentary · 2020
Las Cintas de la Alegría
The Joy Tapes. Official selection, 9th FIDBA (Buenos Aires). First Film category.
LinkedIn ↗ Email ↗
Writing

A philosopher who thinks
in public

Diagnostics, analysis, and questions on AI, humanities, and the university as institution. No hot takes — thinking with sources and theses.

March 2026 Manifesto
Fluency is not knowledge

Ten declarations on verification, abstention, and documentary reliability in the age of AI. With Borges, Kafka, Foucault, and Picasso.

→
March 2026 Column
I, Claudius nº 1 — Four Machines Arguing Whether a Machine Can Read

On an accidental symposium, a question about the author, and a formula about love.

→
March 2026 Manifesto
Four Theses on AI and Judgment

The problem with artificial intelligence in knowledge domains is not a problem of power. It is a problem of judgment. Four pieces, one argument.

→
Marzo 2026 Essay
Silence as Technical Capacity

Language models don't know how to stay silent. When they don't know, they talk more. Can deliberate silence — the kind born of judgment and not of ignorance — be instructed in an AI system?

→
Marzo 2026 Essay
The Question Has Changed

The first generation of digital humanities solved access. The second faces the harder problem: judgment. The leap is not technical — it is epistemological.

→
Marzo 2026 Reflection
Bolaño, Clairvoyance, and Artificial Intelligence

Language models master correctness. But what makes creative work valuable is not the most probable, but the most precise. And often the most precise thing is what one chooses not to include.

→
Marzo 2026 Diagnostic
From MOOC to AI Tutor: The Question Universities Are Not Yet Asking

In one week, two tools appeared that turn any topic into personalized learning for less than two dollars. The obvious reaction is to ask how to compete with them. The right question is different.

→

Do you work with
complex corpora?

I am interested in speaking with people working in research, libraries, cultural heritage, or AI applied to domains where precision matters.

Send an email → LinkedIn
[email protected]