Seven peer-reviewed publications, four as first author, and one first-author manuscript under review. Also listed on Google Scholar and ORCID.
Best Paper Award
PAUL anchors per-class autoencoders with one to five expert-selected exemplars and expands pseudo-labels over an unlabelled pool, reaching 66.9% accuracy on EMNIST Balanced against 53.3% for the strongest baseline.
FormWR, a 387k-page writer-retrieval benchmark of sparse handwriting on printed forms, and X-VLAD, a learned set-aggregation module reaching 78.8% mAP on HisFragIR20 from a 256-dimensional embedding.
A survey of how modern document analysis pipelines turn historical archival records into structured, searchable information.
If conversation with a language model has therapeutic effects in elder care, what follows if we take the drug metaphor seriously — dosage, side effects, contraindications and informed consent?
Nine commercial and open-weight LVLMs benchmarked for handwriting recognition across four scripts. Feeding whole pages beats line segmentation on every handwritten script tested, and costs less.
Under review
An extended study of how input granularity — page, paragraph or line — shapes what large vision-language models can read.
Removing identifying information from wearable IMU time series with generative adversarial networks, while keeping the data useful for downstream human activity recognition.
Simulation software for projecting the storage and computing resource requirements of future LHC experiments, developed at CERN for the ALICE experiment.