Juan é cofundador e CEO da anyformat. É engenheiro nuclear e doutorado em Física experimental pela Sorbonne Université (Paris 6). Passou anos a construir sistemas de IA em produção na Clarity AI e foi o primeiro engenheiro de GenAI da Latency. Escreve sobre inteligência documental, arquiteturas de extração agênticas, calibração de confiança e a distância entre as demos de document AI e a produção. A anyformat é apoiada pela Kibo Ventures, 4Founders Capital e Abac Nest Ventures, e trabalha com clientes como L'Oréal, IAG, Iberia e o Governo de Singapura.
Artigos
The Demo Works. Production Is the Benchmark.
· We tested frontier models and dedicated document-AI engines on 1,000+ real documents across four studies: parsing quality, long-document extraction, complex layouts and confidence calibration. anyformat tops every study, and is the only system that pairs frontier-level accuracy with visual citations and calibrated confidence.Long Documents Are the Production Case: Why 300-Page PDFs Break Extraction Systems and How We Solved It
· Most extraction tools demo on 5-page invoices. Production runs on 300-page filings. We explain why long documents break LLMs and chunking pipelines, how the rest of the field is approaching the problem, and the parse-extract architecture anyformat ships so document teams stop firefighting PDFs.If You Can't Point to It, You Can't Trust It: Why Visual Grounding Is the Foundation of Auditable Document AI
· Most document AI systems can't show where extracted values came from. Learn why visual grounding — linking every output to its exact source region — is the key to auditable, trustworthy document automation.Beyond Accuracy: The Document AI Metrics That Actually Predict Production Success
· Accuracy benchmarks hide silent failures in document processing. Learn the 5 metrics — including confidence calibration, straight-through processing rate, and silent failure rate — that separate production-grade IDP systems from demo-ware.The Paper Paradox: Why Document AI Still Hasn't Replaced Manual Work
· 61% of document processing workflows still involve paper. 66% of new projects replace failed ones. The problem isn't the AI. It's trust.The End of 'We'll Build It In-House': 5 Document Processing Predictions for 2026
· Why this is the year enterprises stop reinventing the wheel on document infrastructure. Buy vs. build finally tips—for non-core problems.Making AI Data Extractions Trustworthy
· This piece introduces a method for scoring the confidence of AI-generated structured outputs, like JSONCómo desbloquear el valor de los datos no estructurados
· Las empresas acumulan datos sin usar. La IA Generativa convierte ese caos en innovación, eficiencia y ventaja competitiva.Una Nueva Era: Los Nobel de Hopfield, Hinton y Hassabis y el Futuro de la Inteligencia Híbrida
· Los Nobel de Física y Química 2024 reconocen el impacto histórico de la IA en la ciencia y la industria, inaugurando una era de colaboración humano-máquina.
