Yvonne Lim Wenqian

Developed end-to-end deep learning pipelines for automated bird sound classification using PyTorch on BirdSet benchmark datasets.
Processed 150+ hours of audio data across datasets ranging from HSN (5,460 samples, 21 species) to XCM (89,798 samples, 409 species) for multi-label audio classification.
Fine-tuned Bird-MAE Vision Transformer (86M parameters) with Prototypical Part Networks, achieving AUROC of 0.91 and mAP of 56% on POW dataset (48 species).
Conducted comparative experiments evaluating multiple Bird-MAE architectural variants (Base, Large, Huge) using prototypical probing methods on distributed NVIDIA A100 GPU infrastructure.
Optimised model performance using AUROC, mAP, and multi-label classification metrics.

Developed end-to-end deep learning pipelines for automated bird sound classification using PyTorch on BirdSet benchmark datasets, processing 150+ hours of audio data across datasets ranging from HSN (5,460 samples, 21 species) to XCM (89,798 samples, 409 species) for multi-label audio classification, Fine-tuned Bird-MAE Vision Transformer (86M parameters) with Prototypical Part Networks, achieving AUROC of 0.91 and mAP of 56% on POW dataset (48 species), Conducted comparative experiments evaluating multiple Bird-MAE architectural variants (Base, Large, Huge) using prototypical probing methods on distributed NVIDIA A100 GPU infrastructure, optimising model performance using AUROC, mAP, multi-label classification metrics

Built an end-to-end ML study scheduler for predictive study optimization.
Trained more than 7 models to analyze student data.
Built a Knowledge Graph visualization engine mapping relationships between courses.
Presented at AI For Education (AIFE) 2025 Conference (NTU-NVIDIA).
Implemented geospatial filtering and distance-based recommendations of 8+ Singapore location datasets.
Built GPS-verified check-in system with JWT authentication and RESTful APIs, achieving under 2 seconds response time.

Built an end-to-end ML study scheduler for predictive study optimization, training >7 models to analyze student data and a Knowledge Graph visualization engine mapping relationships between courses, Presented at AI For Education (AIFE) 2025 Conference (NTU-NVIDIA), Implemented geospatial filtering and distance-based recommendations of 8+ Singapore location datasets, Built GPS-verified check-in system with JWT authentication and RESTful APIs, achieving

Contributed to research on multimodal recommender systems under senior researchers.
Implemented components for a causal inference framework (MGCE+) that addresses bias in recommendations.
Implemented attention-based neural network components for combining multiple data types (images, text) in recommendation systems.
Worked with PyTorch on Amazon datasets (50K+ user-item interactions).
Assisted with experimental validation across three Amazon benchmark datasets.
Ran experiments and analyzed results that showed improvements in standard recommendation metrics (Recall@20, NDCG@20).
Built automated RML mapping generator using Python and pandas to transform educational datasets into RDF knowledge graphs.
Enabled semantic relationships between 500+ course topics and learning objectives.
Engineered data pipeline to parse Excel workbooks, normalize schema inconsistencies, export to CSV format, and generate Turtle (.ttl) RML mappings for processing with RMLMapper to produce RDF triples.
Designed schema mapping curriculum hierarchy, student-module enrollment data, and Bloom's level assessments into linked RDF resources, facilitating graph-based querying and visualization of learning pathways in NTU's production system.

Contributed to research on multimodal recommender systems under senior researchers, implementing components for a causal inference framework (MGCE+) that addresses bias in recommendations, Implemented attention-based neural network components for combining multiple data types (images, text) in recommendation systems, working with PyTorch on Amazon datasets (50K+ user-item interactions), Assisted with experimental validation across three Amazon benchmark datasets, running experiments and analyzing results that showed improvements in standard recommendation metrics (Recall@20, NDCG@20), Built automated RML mapping generator using Python and pandas to transform educational datasets (curriculum nodes, module prerequisites, student assessments, Bloom's taxonomy levels) into RDF knowledge graphs, enabling semantic relationships between 500+ course topics and learning objectives, Engineered data pipeline to parse Excel workbooks, normalize schema inconsistencies, export to CSV format, generate Turtle (.ttl) RML mappings for processing with RMLMapper to produce RDF triples, Designed schema mapping curriculum hierarchy (prerequisite/corequisite/subtopic relationships), student-module enrollment data, and Bloom's level assessments into linked RDF resources, facilitating graph-based querying and visualization of learning pathways in NTU's production system