Vijay Pahuja and Vishal Padh, Technical Debt, AI, Software Engineering, Development Tools United States of America
Technical Debt is one of the biggest issues hindering the digital transformation of organizations. Cost of addressing debt has been rising. AI powered tools can overcome problems of traditional tools as they continuously learn and adapt new patterns. They can proactively detect issues, suggest refactoring, and provide insight to areas of improvement in the codebase, pushing for more sustainable software development practices. While AI offers tremendous potential for managing and reducing technical debt, AI based tools come with their own challenges as AI is heavily dependent on the quality and quantity of data on which they are trained. As organizations rely more and more on AI, they may end up with monotonous codebases producing mediocre products as use of AI will lead to skill degradation and affect critical thinking.
Technical Debt, AI, Software Engineering, Development Tools.
Paris Vuong1, Moddwyn Andaya2, 1Portola High School, 1001 Cadence, Irvine, CA 92618, 2California State Polytechnic University, Pomona, CA, 91768
The issue of limited cultural exposure is present in society due to fast-paced lifestyles and lack of awareness. In an effort to solve this problem, I focused on the food aspect of culture and developed a French cooking simulation game that promotes cultural curiosity through interactive gameplay and historical learning [14]. The game includes time-bound cooking challenges with a recipe book that has educational background information on traditional French dishes. Key systems include the cooking mechanics, the recipe-based learning component, and a progression system to maximize engagement. Design challenges included defining player interaction and ensuring cultural content was informative and accessible. A small-scale survey experiment showed high engagement (average enjoyment score: 4.2) and moderate cultural learning (3.7), indicating the game is both entertaining and educational [15]. Compared to other cultural learning tools, this game offers greater accessibility and balance, also providing a scalable way to introduce users to diverse cultures.
Cultural Education, Cooking Simulation, Interactive Learning, French Cuisine.
Ji-Jer Huang Shang-Min Zheng, Southern Taiwan University, Taiwan
The proposed non-image-based Tangible User Interface (TUI) system integrates physical 3D bricks with interactive cognitive Serious Games (SGs). The system identifies both the characteristics and placement of the bricks by reading encoded data stored within embedded chips incorporated into each brick. Currently, the system offers two interactive cognitive games: a memory game and a matching game. A total of 14 elderly individuals participated in this study, comprising 5 cognitively normal elderly participants and 9 individuals diagnosed with mild dementia. All participants completed a two-week intervention that involved playing the memory game. The results demonstrated a statistically significant difference in game performance between the two groups before and after the intervention. These findings suggest that the developed interactive gaming system holds promise in enhancing cognitive function among elderly individuals with Mild Cognitive Impairment (MCI). Furthermore, system usability was assessed using the System Usability Scale (SUS), yielding a mean SUS score of 84.28. This score reflects a high level of user satisfaction, indicating that the proposed interactive games are both accessible and well-received by the target population.
Non-Image-Based 3D brick, Tangible user interfaces, Serious games.
Xia Li, Tanvi Mistry, The Department of Software Engineering and Game Design and Development, Kennesaw State University, Marietta, USA
Version control systems (VCS) play a crucial role by enabling developers to record changes, revert to previous versions, and coordinate work across distributed teams. In version control systems (e.g., GitHub), commit message serves as concise descriptions of code changes made during development. In our study, we evaluate the performance of multi-label commit message classification using p-tuning (learnable prompt templates) through three pre-trained models such as BERT, RoBERTa and DistilBERT. The experimental results demonstrate that RoBERTa model outperforms other two models in terms of the widely used evaluation metrics (e.g., achieving 81.99% F1 score).
Multi-label commit message classification, p-tuning, pre-trained models.
Preety Joshi1 and Shahid Ali2, 1The Testing Consultancy Software Assurance Company, Auckland, New Zealand, 2Department of Information Technology, AGI Institute, Auckland, New Zealand
Testing is considered a key element in any business, especially in product development. Similarly, software development testing plays a key role in ensuring the quality of products. The challenge to perform manual testing is enormous as it takes huge time and effort, especially in rapidly development environments. To tackle this challenge tech experts, deploy automation testing. The focus of this research is to test web-based application via automation scripts. This research presents a proposed solution for Student Management System Test Automation Solution (SMS TAS). The solution is developed to test the functionality of Student Management System automatically. The institute SMS was a web-based application and was developed to manage students, staff, enrolments and other related tasks.
Student Management System, Test Automation Solution, Application Programming Interface, Page Object Model, Cucumber.
Nuwan Kaluarachchi, Arathi Arakala, Sevvandi Kandanaarachchi and Kristen Moore, Technical University of Crete, , Greece
This study proposes a multi-objective Mayfly Optimization Algorithm to tackle the NP-hard problem of selecting tuning hyperparameters and features in Random Forest models. The optimization simultaneously targets enhanced accuracy and F1 score, while minimizing the number of input variables, thus improving model efficiency without compromising performance. The results indicate that the optimized model performs better than standard configurations, achieving notable improvements in both predictive metrics and model simplicity. A critical aspect of the study lies in the interpretability of the model’s outputs through SHAP (SHapley Additive exPlanations) values, which offer transparency into how individual features influence predictions. In the context of large-scale customer datasets, the use of SHAP proved valuable in isolating the dominant predictors of churn, such as age, number of products, and activity status. The SHAP beeswarm plot demonstrates that older age and fewer products strongly correlate with higher churn risk, whereas active users and higher engagement tend to reduce it. While some features like salary and credit score were less impactful, the explainable outputs enhanced the trust and usability of the model, especially in complex environments with high-dimensional data. The study underscores the value of combining optimization with explainability for handling real-world big data applications, where accuracy alone isnt enough without knowing why a model behaves as it does.
random forest, mayfly, hyperparameters, feature selection, SHAP .
Archit Yajnik Department of Mathematics, Sikkim Manipal Institute of Technology, Sikkim, India
The article represents the Grammar corrector (GC) based on the syntactic and semantic information of a Nepali sentence. Skip-gram model is used for the word to vector encoding. Window size of 3 context words is employed for the word to vector encoding. The network is trained up to the negative log entropy goes to 0.05. The network is tested over 500 incorrect syntactics and semantics of Nepali sentences. The network has suggested the corrections with the accuracy of 96.4%.
Skip-Gram, Grammar Corrector, Word Embedding.
Michael O. Eniolade Department of Information Technology, University of the Cumberlands, USA
Machine learning (ML) represents a pivotal development in the technological era, offering solutions across numerous industries by enabling systems to learn and adapt without explicit programming. The expansive field of ML applications has significantly influenced healthcare, finance, natural language processing (NLP), and computer vision, among others. This paper explores the fundamental principles of ML applications, highlighting key developments, use cases, challenges, and emerging trends. Drawing upon recent scholarly research and industry practices, we illustrate how ML has evolved from theoretical concepts to practical tools that transform industries. The findings suggest that while ML offers unprecedented opportunities for innovation, it also presents significant challenges that demand ethical consideration and methodological advancement.
Machine Learning (ML), applications, natural language processing (NLP), healthcare, finance, digital, computer vision
Alejandro Alvarez Castro1 and Joaqu´ın Ordieres-Mere2, 1Master student at the AI master. Universidad Politecnica de Madrid. Madrid, 28040, 2Industrial Engineering School. Universidad Politecnica de Madrid. Madrid, 28006
Earnings calls represent a uniquely rich and semi-structured source of financial communication, blending scripted managerial commentary with unscripted analyst dialogue. While recent advances in financial sentiment analysis have integrated multimodal signals, such as textual content and vocal tone—most systems rely on flat document-level or sentence-level models, failing to capture the layered discourse structure of these interactions. This paper introduces a novel multimodal framework that encodes entire earnings calls as hierarchical discourse trees. Each node, comprising either a monologue or a question–answer pair, is enriched with emotional signals derived from text, audio, and video, as well as structured metadata including coherence scores, topic labels, and answer coverage assessments. A twostage transformer architecture is proposed: the first encodes multimodal content and discourse metadata at the node level using contrastive learning, while the second synthesizes a global embedding for the entire conference. Experimental results reveal that the resulting embeddings form stable, semantically meaningful representations that reflect affective tone, structural logic, and thematic alignment. Beyond financial reporting, the proposed system generalizes to other unscripted, high-stakes communicative domains such as telemedicine, education, and political discourse, offering a robust and explainable approach to multimodal discourse representation.
Multimodal Learning, Neural Machine Translation (NMT), Speech-Text Alignment, Crossmodal Embeddings, Transformer Models, Multilingual Corpora, Representation Learning, Sequence-toSequence Models, Self-supervised Learning.
Warda Tariq, Department of Computer Science, Higher School of Economics (HSE University), Moscow, Russia
This paper presents a reproducible and systematically improved pipeline for Bartangi language lemmatization and word embedding modeling, designed to advance the state of computational methods for this under-resourced language. Building upon prior work, we address several key challenges that limited the ef ectiveness of earlier versions. Specifically, we focus on the issues of over-filtering, which resulted in the exclusion of important lexical items; af ix misanalysis, which introduced incorrect lemma representations; and sentence-level sparsity, which reduced the contextual richness of the lemmatized data. These problems significantly impacted the quality, interpretability, and usability of the resulting corpora in linguistic and computational tasks.To resolve these challenges, we propose a set of refined morphological parsing and part-of-speech (POS) filtering rules. Our approach ensures that semantically and syntactically meaningful tokens, including infinitive verbs, common nouns, and core grammatical elements, are preserved during preprocessing. At the same time, af ix-like tokens, short or malformed lemmas, and other non-essential forms are systematically removed. This process produces a clean and linguistically reliable Bartangi corpus, which is well-suited for training distributional semantic models.Using this corpus, we train two word embedding models based on the Word2Vec framework, namely Skip-gram and CBOW. These models capture the co-occurrence patterns and latent structure of Bartangi words in dif erent contexts. To better understand and compare the learned representations, we employ dimensionality reduction techniques, including Principal Component Analysis (PCA) and t-distributed Stochastic Neighbor Embedding (t-SNE). These visualizations reveal distinct clustering behaviors, where Skip-gram embeddings more ef ectively capture rare and context-sensitive words, while CBOW embeddings are more suitable for modeling frequent and contextually stable lexical relations. The entire pipeline encompassing the enhanced lemmatization process, the cleaned corpus, trained Word2Vec models, and the visualization tools is publicly released through a dedicated GitHub repository. By making these resources openly available, this work contributes to the growing body of reproducible research in low-resource language processing. Furthermore, it of ers a valuable foundation for future ef orts in endangered language documentation, computational linguistics, and NLP applications targeting Bartangi and related languages.
Bartangi language, Low-resource languages, Lemmatization, Morphological parsing, Word embeddings, Word2Vec (CBOW and Skip-gram), t-SNE visualization, NLP pipeline.
Annie Cui1, Han Tun (Henry) Oo2, 1Orange County School of the Arts, 1010 N Main St, Santa Ana, CA 92701, 2California State Polytechnic University, Pomona, CA, 91768
This research project investigates the effectiveness of sentiment analysis tools on tweets discussing neurodivergent individuals, particularly those with autism. Traditional models like TextBlob often lack the contextual awareness needed to interpret subtle or emotionally complex content. To address this, we developed a system comparing TextBlob and GPT-4 using both classification and regression-based evaluation [1]. A dataset of 100 tweets was analyzed. In the first experiment, GPT-4 achieved a macro F1-score of 0.61, outperforming TextBlob’s 0.58, with both models reaching 62% accuracy. In the second experiment, which evaluated polarity scoring, GPT-4 achieved a MAE of 0.604, RMSE of 0.766, and a correlation of 0.479, compared to TextBlob’s MAE of 0.650, RMSE of 0.778, and correlation of 0.394. These results confirm that GPT-4 provides more accurate and context-sensitive sentiment predictions [2]. This system improves upon prior lexicon-only approaches by combining classification and polarity scoring to offer a comprehensive, real-world analysis of sentiment in neurodivergent conversations.
Sentiment analysis, GPT-4, TextBlob, Autism, Neurodivergent discourse, Twitter data, Natural language processing, Polarity score, Classification metrics, Context-aware AI.
Daniel Zhang1, Jonathan Thamrun2, 1Beckman High School, 3588 Bryan Ave, Irvine, CA 92602, 2California State Polytechnic University, Pomona, CA, 91768
Running is a popular form of exercise across the world with primal origins and plenty of health benefits. However, running injuries are very common and pose a major threat to competitive and even casual athletes. Traditionally, runners seek medical experts like physiotherapists to treat their ailments. However, this can often be expensive, timeconsuming, and have lengthy wait times for the patient. To address this, we created an AI-driven solutionthat utilizes Natural Language Processing (NLP) and Computer Vision to diagnose users and recommend treatments. Although our method isn’t meant to replace medical experts, it can act as a swift, mobile first opinion for injuredrunners. Our solution comes in the form of a simple mobile application. The user can fill out multiple optional sections including a pain questionnaire, image upload, and additional information section to receive AI generatedrecommendations. This application was built using the Flutter framework and leverages Firebase for data storage. Additionally, it uses the OpenAI API to send requests to our AI model (gpt 4.1) and receive its responses. We facedavariety of challenges while tailoring the prompt and designing the UI to give the user the best results in the most readable manner. We addressed these challenges through iterative development and frequent testing. Inexperimentation, we found that the model could detect injuries in knee X-Rays with 64% accuracy and generateresponses in around 82 seconds on average. Overall, the model has potential to be used as a first-aid kid of sorts, giving the user an initial idea of what they may be injured with and how they could treat it. It is easy-to-use andcanbe accessed anywhere as long as the user has their phone.
Machine Learning, Computer Vision, Nature Language Processing.
Haiyi Yang1, Ang Li2, 1Capital Normal University High School, Beiwa Street No.33, Haidian District, Beijing, China, 2California State Polytechnic University, Pomona, CA, 91768
This paper presents a mobile application designed to support mental health through AI-generated art, music, and journaling, guided by emotion detection. The app uses facial emotion recognition and journal sentiment to generate daily personalized images and music that promote emotional healing. Built with Flutter, Python, OpenAI models, and Supabase, the system integrates real-time chat, media storage, and adaptive content generation [1]. Three key systems—emotion detection, AI art, and AI music—are evaluated through experiments and compared with scholarly research. The FER model achieved 70 percent accuracy, and DALL·E-generated images scored highly in emotional alignment [2]. Compared to similar projects, our app stands out for its simplicity, accessibility, and user-driven personalization. Limitations include the need for multimodal input and further validation in clinical settings. Future improvements will focus on enhancing emotional accuracy and user safety. Overall, the project demonstrates a promising and scalable approach to digital art therapy and emotional support.
Mental health, Emotion detection, Art therapy, AI-generated music, Mobile application.
Hadi Zaatiti1, Hatem Hajri2, Osama Abdullah1, and Nader Masmoudi3,4, 1Bio-Medical Imaging Core, Core Technology Platforms, New York University Abu Dhabi, PO Box 129188, Abu Dhabi, UAE, 2Institut de recherche technologique SystemX, 8 Avenue de la Vauve, 91120 Palaiseau, France, 3NYUAD Research Institute, New York University Abu Dhabi, PO Box 129188, Abu Dhabi, UAE, 4Courant Institute of Mathematical Sciences, New York University, 251 Mercer Street, New York, NY 10012, USA
Modern Arabic ASR systems such as wav2vec 2.0 excel at word- and sentence-level transcription, yet struggle to classify isolated letters. In this study, we show that this phoneme-level task, crucial for language learning, speech therapy, and phonetic research, is challenging because isolated letters lack co-articulatory cues, provide no lexical context, and last only a few hundred milliseconds. Recogniser systems must therefore rely solely on variable acoustic cues, a difficulty heightened by Arabic’s emphatic (pharyngealized) consonants and other sounds with no close analogues in many languages. This study introduces a diverse, diacritised corpus of isolated Arabic letters and demonstrate that state-of-the-art wav2vec 2.0 models achieve only 35 % accuracy on it. Training a lightweight neural network on wav2vec embeddings raises performance to 65 %. However, adding a small amplitude perturbation (ϵ = 0.05) cuts accuracy to 32 %. To restore robustness, we apply adversarial training, limiting the noisy-speech drop to 9 % while preserving clean-speech accuracy. We detail the corpus, training pipeline, and evaluation protocol, and release, on-demand, data and code for reproducibility. Finally, we outline future work extending these methods to word- and sentence-level frameworks, where precise letter pronunciation remains critical.
Arabic letters pronunciation, adversarial training, classification
Aman Sharma, Vellore Institute of Technology, India
We investigate dual scaling laws governing arithmetic performance in large language models (LLMs): how performance scales with problem complexity (digit count) and context window size. Our results reveal a ”Context Constraint Effect” where reasoning-based prompting strategies significantly underperform direct answer strategies when context is limited. Through experiments across multiple models (Claude-3.7- Sonnet, GPT-4o, DeepSeek Chat), we demonstrate that performance decline follows mathematical patterns described by power law (R² = 0.990) and exponential (R² = 0.989) models. We introduce a framework predicting minimum context size required for different problem complexities and identify strategy crossover points where reasoning approaches outperform direct answers.
Large Language Models, Arithmetic Reasoning, Scaling Laws, Context Windows, Chain-of-Thought.
Salih Mansur, Harvard University, Division of Continuing Education, Cambridge, USA
This practice-based study examines the impact of an AI-supported digital literacy course (DigiLit) on adult learners’ technology skills and confidence. Using a mixed-methods approach, we collected before-and-after surveys and written reflections from 11 participants, primarily from service-based professions. Learners used ChatGPT and Google Workspace tools to complete real-life tasks. Results showed a 40% increase in confidence using AI tools and marked improvement in digital fluency. Qualitative data highlighted themes such as overcoming fear of technology and applying new skills in work scenarios. The findings suggest that accessible, well-structured digital literacy programs can significantly support adult learners navigating modern digital environments.
artificial intelligence (AI), digital literacy, adult education (Andragogy), Google tools, technology confidenc.
Jiangnan Liu1, Justin Dang2, 1Fairmont Preparatory Academy, 2200 W Sequoia Ave, Anaheim, CA 92801, 2California State Polytechnic University, Pomona, CA, 91768
Manual rocket flight and landing systems are prone to human error and can be quite costly and resource intensive [1]. In order to advance properly into space travel, we should try and devise methods to make these systems safe and autonomous. Our project idea was centered around creating an autonomous rocket landing mechanism using reinforcement learning simulations that can withstand various different environmental conditions. This system uses the Unity engine as well as a PPO reinforcement learning algorithm in order to train an agent to optimize a rocket landing algorithm based on trial, error, and rewards [2]. We overcame several challenges that came with implementing this system, including optimizing the reward function, and making sure the simulation could reliably model the real world as well as possible. To test this system’s efficiency and accuracy, we devised many different scenarios involving different conditions regarding rocket flight, angle, and wind speed, and found that the model could under various conditions consistently handle these different conditions and safely land.
Autonomous Landing, Reinforcement Learning, Rocket Simulation, PPO Algorithm
