This paper presents an enterprise HR FAQ chatbot built using Retrieval-Augmented Generation (RAG) to provide grounded and verifiable responses to employee queries. To address domain mismatch in the Doc2Dial dataset, a custom enterprise HR knowledge base was constructed. The system uses FAISS for dense vector retrieval and IBM Granite-3.1-2B-Instruct for response generation, deployed via a Databricks pipeline and a Hugging Face Spaces interface. The system was evaluated through multi-model benchmarking against Phi-3.5-Mini, Qwen-2.5-1.5B, and SmolLM-1.7B under CPU constraints. Granite achieved the best balance between grounding accuracy and citation compliance, with a latency of 8.5 seconds and a perplexity of 5.09. In contrast, Phi-3.5-Mini achieved lower perplexity (3.35) but higher latency (12.2 seconds), while smaller models showed reduced grounding quality and inconsistent citation behavior. A strict L2 distance threshold (> 1.5) was implemented to prevent hallucinations, ensuring deterministic fallback responses for out-of-scope queries. Results demonstrate that a lightweight, constraint-aware RAG architecture can deliver reliable, citation-grounded answers in enterprise environments.
This paper investigates lightweight adaptation of a conversational large language model for topical open-domain dialogue. Using the Topical-Chat dataset, conversations were converted into supervised next-response prediction examples, where the model learns to generate the next assistant reply from preceding dialogue history. A Qwen2.5-1.5B-Instruct model was adapted using Low-Rank Adaptation (LoRA) under Kaggle GPU constraints. In addition, a local retrieval-augmented generation (RAG) baseline was implemented using wiki-backed Topical-Chat factual sections and BM25 retrieval. Fine-tuning experiments across progressively larger training subsets show that validation loss decreased from 3.1550 in a 300-sample pilot run to 2.6430 in the strongest reported run, with perplexity decreasing to 14.0557. The RAG baseline provided explicit evidence retrieval for factual prompts, while LoRA fine-tuning improved dialogue modeling. At this stage ,the system is suitable for supervised prototyping rather than direct public deployment.
Khaled Toffaha, Ahmed Saad, Michael A. Kortt and Samer
Ellahham
Presenting author: Ahmed Saad
Abstract
The rapid digital transformation and economic diversification taking place across the Gulf Cooperation Council (GCC) countries have been associated with significant changes in labor market dynamics. This study, using the United Arab Emirates (UAE) as a case example, has conducted a detailed analysis of shifting skill requirements. A mixed-methods approach was employed, combining quantitative analysis of over 6,000 job advertisements collected from LinkedIn between 2018 and 2024 across the engineering, healthcare, and finance and marketing sectors. The job descriptions were first processed using OpenAI's extraction tools, and then Latent Dirichlet Allocation (LDA) was applied to identify major skill clusters and track their changes over time. A comparative analysis was conducted to examine the trends in required skills over different years. A 25% increase in IT-related job postings was observed compared to earlier years, and more defined clusters related to programming and data analysis were detected in the recent data. In contrast, broader and less specific skill categories were found in the earlier listings. It was also found that Data Scientist, AI Engineer, and Digital Marketing Specialist were the most frequently advertised roles, together making up 30% of the postings. A clear shift was identified from job roles that mainly required soft skills to those that combine technical and interpersonal skills. Based on the results, a stronger focus on technical training in education programs is recommended, while continuing to support the development of soft skills, to better prepare the workforce for the evolving demands of the GCC job market.
Sara Bahlooq, Sara Zaidan, Hind Abdulla, Dalal Almansoori, Ahmad Mohamad and Firda Rahmadani
Presenting author: Sara Bahlooq
Abstract
Surgical site infections (SSIs) remain a major healthcare challenge, affecting patient outcomes and increasing operational costs. Although artificial intelligence (AI) shows potential for improving the prediction, detection, and prevention of SSIs, the literature lacks a comprehensive systems design framework to support its implementation. This study addresses this gap by developing an enhanced systems-thinking, AI-driven decision support framework that integrates design-thinking principles. The framework comprises four components: people, systems, design, and risks. Through a review of existing literature and current practices, the study identifies key factors influencing AI implementation in SSI management, including user-centered design, system interoperability, iterative prototyping, and risk assessment. The resulting framework provides a structured approach for integrating AI into SSI management and offers guidance for researchers, clinicians, and healthcare administrators. The study concludes with recommendations for applying the framework in healthcare settings, emphasizing interdisciplinary collaboration, stakeholder engagement, and proactive risk management to improve patient safety and reduce healthcare-associated infections.
The consistency of production outputs relies on the operational effectiveness of manufacturing resources and the reliability of interconnected subsections of the production lines and equipment. In this study, lower production throughput, linked to operational inefficiencies and a higher frequency of longer downtimes in automated brick-making machinery, is observed in factory “A”. These random failures of automated brick manufacturing machinery at the factory where the research was conducted resulted in an average annual production loss of 13% in 2023, based on figures extracted from Network and Equipment Performance System Quality Assurance (NEPSQA), herein referred to as factory "A" for anonymity. Based on NEPSQA-tracked estimates, over 85% of these production losses are linked to downtime caused by inefficiencies in the production line. Lower production line reliability is caused by a higher number of deficiencies triggering manufacturing system failures. An Artificial Neural Network (ANN) algorithm was applied in this research to identify and distinguish production line deficiencies. The approach facilitated the development of appropriate Reliability-Centred Maintenance (RCM) activities to reduce failure occurrences. The ANN algorithm improved the RCM by reducing the Average Hours Interruption Indices (AHII) by 56.9% and the Average Frequency Interruption Indices (AFII) by 25.34% per month, as reported in the NEPSQA data.
