Selected work

Representative Papers

A Modular Part-of-Speech Tagger for Scottish Gaelic using spaCy

Aug 05, 2026

This study addresses part-of-speech tagging for Scottish Gaelic—a low-resource, morphologically complex language—under conditions of extreme data scarcity and without access to external linguistic resources. Building upon the spaCy framework, the authors develop a lightweight, modular supervised tagger that relies solely on a reference corpus of Scottish Gaelic, training separate models for fine-grained and coarse-grained tagsets without leveraging pretrained language models or external embeddings. Experimental results demonstrate that the fine-grained and coarse-grained models achieve accuracies of 88.6% and 93.7%, respectively, matching the performance of current state-of-the-art systems. These findings underscore the effectiveness and practicality of off-the-shelf NLP pipelines for tackling morphologically rich, low-resource languages when appropriately adapted within a supervised learning paradigm.

0 citationsRead paper

Large Language Models for Software Engineering Diagrams: A Systematic Review of UML and ER modelling

Jul 28, 2026

This study addresses the lack of a systematic review on the application of large language models (LLMs) in UML and ER diagram modeling, where inconsistent evaluation, uneven coverage, and poor reproducibility persist. Through a systematic literature review of 64 studies published between 2023 and 2025, it examines the use of GPT-family models in tasks involving natural language–to–diagram generation, transformation, and validation, analyzing their methodologies, task formulations, and evaluation practices. The findings reveal a predominant focus on class diagram modeling, with significant underrepresentation of behavioral diagrams and data modeling. Common challenges include semantic errors, hallucinations, prompt sensitivity, and heterogeneous evaluation protocols. This work provides the first consolidated synthesis of LLM-driven diagram modeling research, highlighting prevailing trends, critical gaps, and promising directions such as standardized benchmarks and multi-view consistency.

0 citationsRead paper

SYNAPSE: A Multi-LLM Orchestrated AI Tutor for Secure Software Development Education with Neurodivergent-First Design

Jul 16, 2026

This study addresses the persistent lag of secure programming education behind foundational coding instruction and its inadequate support for neurodiverse learners, such as those with ADHD. To bridge this gap, the authors propose an adaptive AI tutoring platform designed explicitly for neurodiverse individuals, introducing a “neurodiversity-first” pedagogical framework. The system integrates 18 consistently available accessibility features and leverages the Model Context Protocol to orchestrate Claude, GPT-4o, and Gemini large language models. Guided by a Socratic three-stage prompting strategy, the platform facilitates a “detect–understand–remediate” learning cycle within the ShopSecure vulnerability environment. Pilot evaluation demonstrates a System Usability Scale score of 76.4, user engagement of 4.2/5, and comparable cognitive load between neurodiverse and neurotypical learners, affirming the platform’s innovative contributions to accessibility, personalization, and instructional efficacy.

0 citationsRead paper

AI Prototyper: A Figma Plugin for Decomposition-Based GUI Prototyping with LLMs

Jul 16, 2026

This work addresses the time-intensive nature of GUI prototyping and its reliance on specialized expertise, particularly the limited support existing tools offer for efficiently evolving non-code artifacts during iterative requirement changes. The authors propose an automated approach leveraging Retrieval-Augmented Generation (RAG) and human-AI collaboration: it first parses user-provided natural language inputs—supporting Chinese, English, and Thai—to extract GUI design features, then retrieves relevant components from a custom library of 32 base elements to generate editable Figma prototypes with automatic layout. Built on the Gemini 2.5 Flash large language model and a new Node.js/Express architecture, the system incorporates a user-adjustable feature refinement step prior to prototype generation. Preliminary evaluation shows that users produce more prototypes within a fixed timeframe, and expert assessments across nine quality dimensions significantly favor the generated outputs over manually crafted ones.

0 citationsRead paper

What Do AI Agents Actually Change? An Empirical Taxonomy of Mutation Patterns in Performance-Improving Pull Requests

Jul 06, 2026

This study addresses the lack of transparency in code-modification behaviors of current AI-powered coding agents during performance optimization. It presents the first empirical analysis of 1,254 code diffs from 216 AI-generated performance-optimization pull requests, systematically annotated using a dual-LLM cross-validation protocol based on an 18-category syntactic mutation taxonomy derived from Genetic Improvement (GI). The findings reveal a strong preference by AI agents for three mutation types: identifier renaming (37.0%), object creation (26.4%), and type changes (22.7%)—a stark contrast to traditional GI datasets, where over 84% of mutations involve no change. These results suggest that agent identity and optimization strategy can serve as effective priors for narrowing the search space of Search-Based Software Engineering (SBSE) operators.

0 citationsRead paper

Recent publications

Latest Papers

A Modular Part-of-Speech Tagger for Scottish Gaelic using spaCy

Aug 05, 2026

This study addresses part-of-speech tagging for Scottish Gaelic—a low-resource, morphologically complex language—under conditions of extreme data scarcity and without access to external linguistic resources. Building upon the spaCy framework, the authors develop a lightweight, modular supervised tagger that relies solely on a reference corpus of Scottish Gaelic, training separate models for fine-grained and coarse-grained tagsets without leveraging pretrained language models or external embeddings. Experimental results demonstrate that the fine-grained and coarse-grained models achieve accuracies of 88.6% and 93.7%, respectively, matching the performance of current state-of-the-art systems. These findings underscore the effectiveness and practicality of off-the-shelf NLP pipelines for tackling morphologically rich, low-resource languages when appropriately adapted within a supervised learning paradigm.

0 citationsRead paper

Large Language Models for Software Engineering Diagrams: A Systematic Review of UML and ER modelling

Jul 28, 2026

This study addresses the lack of a systematic review on the application of large language models (LLMs) in UML and ER diagram modeling, where inconsistent evaluation, uneven coverage, and poor reproducibility persist. Through a systematic literature review of 64 studies published between 2023 and 2025, it examines the use of GPT-family models in tasks involving natural language–to–diagram generation, transformation, and validation, analyzing their methodologies, task formulations, and evaluation practices. The findings reveal a predominant focus on class diagram modeling, with significant underrepresentation of behavioral diagrams and data modeling. Common challenges include semantic errors, hallucinations, prompt sensitivity, and heterogeneous evaluation protocols. This work provides the first consolidated synthesis of LLM-driven diagram modeling research, highlighting prevailing trends, critical gaps, and promising directions such as standardized benchmarks and multi-view consistency.

0 citationsRead paper

SYNAPSE: A Multi-LLM Orchestrated AI Tutor for Secure Software Development Education with Neurodivergent-First Design

Jul 16, 2026

This study addresses the persistent lag of secure programming education behind foundational coding instruction and its inadequate support for neurodiverse learners, such as those with ADHD. To bridge this gap, the authors propose an adaptive AI tutoring platform designed explicitly for neurodiverse individuals, introducing a “neurodiversity-first” pedagogical framework. The system integrates 18 consistently available accessibility features and leverages the Model Context Protocol to orchestrate Claude, GPT-4o, and Gemini large language models. Guided by a Socratic three-stage prompting strategy, the platform facilitates a “detect–understand–remediate” learning cycle within the ShopSecure vulnerability environment. Pilot evaluation demonstrates a System Usability Scale score of 76.4, user engagement of 4.2/5, and comparable cognitive load between neurodiverse and neurotypical learners, affirming the platform’s innovative contributions to accessibility, personalization, and instructional efficacy.

0 citationsRead paper

AI Prototyper: A Figma Plugin for Decomposition-Based GUI Prototyping with LLMs

Jul 16, 2026

This work addresses the time-intensive nature of GUI prototyping and its reliance on specialized expertise, particularly the limited support existing tools offer for efficiently evolving non-code artifacts during iterative requirement changes. The authors propose an automated approach leveraging Retrieval-Augmented Generation (RAG) and human-AI collaboration: it first parses user-provided natural language inputs—supporting Chinese, English, and Thai—to extract GUI design features, then retrieves relevant components from a custom library of 32 base elements to generate editable Figma prototypes with automatic layout. Built on the Gemini 2.5 Flash large language model and a new Node.js/Express architecture, the system incorporates a user-adjustable feature refinement step prior to prototype generation. Preliminary evaluation shows that users produce more prototypes within a fixed timeframe, and expert assessments across nine quality dimensions significantly favor the generated outputs over manually crafted ones.

0 citationsRead paper

What Do AI Agents Actually Change? An Empirical Taxonomy of Mutation Patterns in Performance-Improving Pull Requests

Jul 06, 2026

This study addresses the lack of transparency in code-modification behaviors of current AI-powered coding agents during performance optimization. It presents the first empirical analysis of 1,254 code diffs from 216 AI-generated performance-optimization pull requests, systematically annotated using a dual-LLM cross-validation protocol based on an 18-category syntactic mutation taxonomy derived from Genetic Improvement (GI). The findings reveal a strong preference by AI agents for three mutation types: identifier renaming (37.0%), object creation (26.4%), and type changes (22.7%)—a stark contrast to traditional GI datasets, where over 84% of mutations involve no change. These results suggest that agent identity and optimization strategy can serve as effective priors for narrowing the search space of Search-Based Software Engineering (SBSE) operators.

0 citationsRead paper