Selected work

Representative Papers

PRomop: A Decision-Ready Longitudinal Patient Health Record on the OMOP Common Data Model

Jul 15, 2026

This study addresses the challenges of fragmented medical data, storage-centric architectures, and redundant reconstruction of clinical states for downstream applications by proposing PRomop—a longitudinal patient health record representation grounded in the OMOP Common Data Model (CDM) version 5.4, extended with oncology-specific elements. PRomop compresses each patient’s complete clinical history into a single-row, 286-column decision-ready format through flat projection and precomputes key clinical states to enable efficient reuse while preserving full compatibility with OMOP standards. Deployed across two oncology institutions, the approach supports eligibility screening for 19,500 clinical trials; empirical evaluation on synthetic data demonstrates a 23.9-fold acceleration in trial matching and a 30- to 200-fold reduction in query complexity.

0 citationsRead paper

Vibe coding before the trend

May 08, 2026

This study investigates how non-traditional programming learners reshape their learning trajectories with the support of generative AI and identifies effective pedagogical strategies for instructor guidance. Through “vibe coding” instructional experiments conducted across multiple universities, combined with reflective practice and qualitative observation, the research systematically documents, for the first time, positive responses from non-technical students toward AI-powered programming tools. Findings reveal that learners widely regard AI as an essential professional skill, particularly valuing its accessibility, and demonstrate a shift in focus from memorizing syntax to developing higher-order thinking and evaluative competencies. The work proposes a novel educational paradigm framing AI as a collaborative learning partner rather than a replacement, thereby affirming the potential and feasibility of generative AI in interdisciplinary programming education.

0 citationsRead paper

Institutional AI Sovereignty Through Gateway Architecture: Implementation Report from Fontys ICT

Dec 04, 2025

Commercial AI subscription tools compromise institutional data sovereignty, entail GDPR compliance risks, and exacerbate service inequities in higher education. This study proposes an AI Sovereignty Gateway Platform tailored for applied universities, featuring a three-tier architecture: (1) an identity-authenticated frontend, (2) a policy-driven gateway core enabling EU-default routing and granular budget control, and (3) a Model Card Provider layer encapsulating both commercial and open-source models to ensure unified model governance and full-chain auditability. The work introduces the first institution-level AI gateway paradigm, elevating AI from a technical utility to a strategic function, and establishes an AI Director role integrating technical, governance, and pedagogical responsibilities. A six-month empirical pilot with 300 users achieved zero privacy incidents, 100% EU-hosted infrastructure, complete model transparency and selection, and high user adoption—validating a sovereign, compliant, controllable, and equitable AI service framework for higher education.

0 citationsRead paper

Benchmarking Energy Efficiency of Large Language Models Using vLLM

Sep 10, 2025

The high energy consumption of large language models (LLMs) severely hinders their sustainable deployment, yet existing energy-efficiency evaluations rely heavily on idealized benchmarks that poorly reflect real-world production workloads. To address this gap, we propose the first energy-efficiency benchmarking framework tailored to realistic LLM inference loads. Built upon vLLM, it establishes a multi-concurrent, dynamically scheduled testbed that emulates production-grade request patterns. We systematically measure and analyze energy consumption across diverse model scales, architectures, and inference workloads. Through cross-model and cross-configuration empirical studies, we quantitatively uncover previously unreported nonlinear relationships between energy efficiency and key factors—including parameter count, attention mechanism design, and hardware utilization. This work not only demonstrates the feasibility of production-relevant energy-efficiency assessment but also delivers a reproducible, extensible quantitative toolkit and actionable optimization guidelines—laying both methodological foundations and practical evidence for green AI systems.

0 citationsRead paper

CrypticBio: A Large Multimodal Dataset for Visually Confusing Biodiversity

May 16, 2025

Automatic identification of cryptic species—visually indistinguishable taxa—is hindered by small, narrow, and manually annotated datasets, limiting scalable biodiversity monitoring. Method: We introduce CrypticBio, the largest public multimodal biodiversity dataset to date, comprising 52K cryptic groups, 67K species, and 166 million images, enriched with scientific nomenclature, multilingual labels, spatiotemporal metadata, and taxonomic hierarchies. We propose the first geographic–temporal dual-dimension modeling framework for crypticity and release CrypticBio-Curate, an open, reproducible data curation pipeline. Contribution/Results: Through multimodal fusion and zero-shot cross-domain evaluation, we demonstrate that geographic context substantially improves recognition accuracy. CrypticBio establishes the first large-scale benchmark targeting endangered, invasive, and undescribed species, enabling field-deployable biological AI and advancing automated biodiversity assessment.

0 citationsRead paper

Recent publications

Latest Papers

PRomop: A Decision-Ready Longitudinal Patient Health Record on the OMOP Common Data Model

Jul 15, 2026

This study addresses the challenges of fragmented medical data, storage-centric architectures, and redundant reconstruction of clinical states for downstream applications by proposing PRomop—a longitudinal patient health record representation grounded in the OMOP Common Data Model (CDM) version 5.4, extended with oncology-specific elements. PRomop compresses each patient’s complete clinical history into a single-row, 286-column decision-ready format through flat projection and precomputes key clinical states to enable efficient reuse while preserving full compatibility with OMOP standards. Deployed across two oncology institutions, the approach supports eligibility screening for 19,500 clinical trials; empirical evaluation on synthetic data demonstrates a 23.9-fold acceleration in trial matching and a 30- to 200-fold reduction in query complexity.

0 citationsRead paper

Vibe coding before the trend

May 08, 2026

This study investigates how non-traditional programming learners reshape their learning trajectories with the support of generative AI and identifies effective pedagogical strategies for instructor guidance. Through “vibe coding” instructional experiments conducted across multiple universities, combined with reflective practice and qualitative observation, the research systematically documents, for the first time, positive responses from non-technical students toward AI-powered programming tools. Findings reveal that learners widely regard AI as an essential professional skill, particularly valuing its accessibility, and demonstrate a shift in focus from memorizing syntax to developing higher-order thinking and evaluative competencies. The work proposes a novel educational paradigm framing AI as a collaborative learning partner rather than a replacement, thereby affirming the potential and feasibility of generative AI in interdisciplinary programming education.

0 citationsRead paper

Institutional AI Sovereignty Through Gateway Architecture: Implementation Report from Fontys ICT

Dec 04, 2025

Commercial AI subscription tools compromise institutional data sovereignty, entail GDPR compliance risks, and exacerbate service inequities in higher education. This study proposes an AI Sovereignty Gateway Platform tailored for applied universities, featuring a three-tier architecture: (1) an identity-authenticated frontend, (2) a policy-driven gateway core enabling EU-default routing and granular budget control, and (3) a Model Card Provider layer encapsulating both commercial and open-source models to ensure unified model governance and full-chain auditability. The work introduces the first institution-level AI gateway paradigm, elevating AI from a technical utility to a strategic function, and establishes an AI Director role integrating technical, governance, and pedagogical responsibilities. A six-month empirical pilot with 300 users achieved zero privacy incidents, 100% EU-hosted infrastructure, complete model transparency and selection, and high user adoption—validating a sovereign, compliant, controllable, and equitable AI service framework for higher education.

0 citationsRead paper

Benchmarking Energy Efficiency of Large Language Models Using vLLM

Sep 10, 2025

The high energy consumption of large language models (LLMs) severely hinders their sustainable deployment, yet existing energy-efficiency evaluations rely heavily on idealized benchmarks that poorly reflect real-world production workloads. To address this gap, we propose the first energy-efficiency benchmarking framework tailored to realistic LLM inference loads. Built upon vLLM, it establishes a multi-concurrent, dynamically scheduled testbed that emulates production-grade request patterns. We systematically measure and analyze energy consumption across diverse model scales, architectures, and inference workloads. Through cross-model and cross-configuration empirical studies, we quantitatively uncover previously unreported nonlinear relationships between energy efficiency and key factors—including parameter count, attention mechanism design, and hardware utilization. This work not only demonstrates the feasibility of production-relevant energy-efficiency assessment but also delivers a reproducible, extensible quantitative toolkit and actionable optimization guidelines—laying both methodological foundations and practical evidence for green AI systems.

0 citationsRead paper

CrypticBio: A Large Multimodal Dataset for Visually Confusing Biodiversity

May 16, 2025

Automatic identification of cryptic species—visually indistinguishable taxa—is hindered by small, narrow, and manually annotated datasets, limiting scalable biodiversity monitoring. Method: We introduce CrypticBio, the largest public multimodal biodiversity dataset to date, comprising 52K cryptic groups, 67K species, and 166 million images, enriched with scientific nomenclature, multilingual labels, spatiotemporal metadata, and taxonomic hierarchies. We propose the first geographic–temporal dual-dimension modeling framework for crypticity and release CrypticBio-Curate, an open, reproducible data curation pipeline. Contribution/Results: Through multimodal fusion and zero-shot cross-domain evaluation, we demonstrate that geographic context substantially improves recognition accuracy. CrypticBio establishes the first large-scale benchmark targeting endangered, invasive, and undescribed species, enabling field-deployable biological AI and advancing automated biodiversity assessment.

0 citationsRead paper