We treat publications, patents, firms, and researcher careers as large-scale relational data and study them with graph neural networks, knowledge graph embeddings, multimodal representation learning, and large language models. Where traditional scientometrics relies on bibliometric indicators and econometric models, we bring modern AI to the same questions: how technologies converge, how capital and talent flow through innovation ecosystems, and how emerging technologies such as AI reshape science and industry.

Methods: graph neural networks, knowledge graph embedding, LLM-based text analysis, contrastive and multimodal learning

Much of the data that matters for innovation, clinical records, transaction logs, production data, cannot leave the organization that owns it. We design federated learning and privacy-preserving AI that let hospitals, banks, and manufacturers build shared models, including generative and foundation models, while keeping data local, with attention to fairness across participants, personalization, and robustness. This line of work connects directly to questions of data governance and data sovereignty in technology policy.

Methods: federated learning, differential privacy, homomorphic encryption, personalized and fair FL, federated generative models

We work with domain partners to solve concrete problems in three flagship areas: 

healthcare (drug recommendation, disease and cost prediction, biological age), 

finance (household financial health, fraud detection, institutional risk), 

manufacturing (multistage process modeling, predictive maintenance). 

Our emphasis is on time-series, graph-structured, and multimodal data, on interpretable and lightweight models including domain-adapted LLMs, and on deployment in real operational settings.