This thesis presents the development of the Welcome Clients platform, a Generative AI–driven solution designed to optimize the planning and management of client visits at Accenture. Unlike traditional manual or semi-automated methods that often fail to address the complexity of personalized visits, the proposed system integrates a GenAI-powered assistant capable of guiding employees through event creation, automatically generating itineraries with precise timestamps, drafting personalized emails for internal teams, and supporting form completion through a conversational chatbot. Unlike priorworks that applied AI mainly in tourism, mobility forecasting, or academic scheduling, this project addresses the enterprise context of corporate hospitality, which demands strict schema adherence, multilingual communication, security compliance, and seamless integration with organizational workflows. The evaluation confirms that the solution significantly reduces preparation time, improves accuracy and efficiency in event planning, and establishes a scalable, replicable model for embedding GenAI in enterprise-level event management.

The paradigm of AI capabilities is reshaping the strategy of consulting firms and unlocking a new frontier of innovation. AI technologies, particularly LLMs, are driving unprecedented change, reinventing consulting practices and enabling efficiency gains never before achieved. However, these new practices require consultants to develop new skills in order to remain competitive in an increasingly dynamic market. Thus, understanding these dynamics required an innovative approach. This study is based on a scoping review (ScoLR) following the framework of Arksey & O’Malley (2005), which allows for mapping existing knowledge, identifying trends, gaps, and implications for the sector. This theoretical framework is complemented by a detailed analysis of the consulting market in Portugal, using the ORBIS (orbis-r1.bvdinfo.com) and base.gov (www.base.gov.pt) databases, with the aim of clarifying the structure of the sector, key players, and the types of services most commonly provided. One of the main contributions of this research is the development of a vulnerability matrix, which classifies consulting services according to their exposure to potential automation or knowledge internalisation by clients, highlighting the areas with the highest long-term strategic value. In parallel, qualitative interviews with industry stakeholders provide valuable insights into how companies are responding to the digital transformation driven by AI, seeking to balance innovation with ethical and operational considerations. The results highlight the transformative potential of AI for productivity and innovation, and offer practical recommendations to help consulting firms make the technological transition and consolidate their competitive advantage.

Building on previous work on incorporating large language mod els (LLM) in gaming, we investigate the possibility of implement ing LLM as evaluating agents of open-ended challenges in serious games and its potential to facilitate a meaningful experience for the player. We contribute with a sustainability game prototype in a single natural language prompt about energy communities and we tested it with 13 participants inside ChatGPT-3.5. Two partici pants were already aware of energy communities before the game, and eight of the remaining 11 gained valuable knowledge about the specifc topic. Comparing ChatGPT-3.5 evaluations of players’ interaction with an expert’s assessment, ChatGPT-3.5 correctly evaluated 81% of player’s answers. Our results are encouraging and show the potential of using LLMs as mediating agents in ed ucational games, while also allowing easy prototyping of games through natural language prompts.

Managing critical infrastructures such as drinking water networks poses a fundamental challenge in transforming vast streams of real-time data into actionable and accessible knowledge. The core operational issue lies in the cognitive burden placed on operators, who must manually correlate numerous disparate variables to answer complex questions, in a process that typically demands years of expert experience. This work addresses the democratisation of expert knowledge by introducing specialised conversational agents based on Large Language Models (LLMs). The proposed approach trains an assistant that learns from domain-specific data, transforming 156 real operational variables from a municipal system into a structured dataset comprising automatically generated technical conversations. The obtained results demonstrate that the specialised conversational agent acts as a senior engineer available 24/7, improving cognitive tasks with key gains: Situational Analysis (+35%), integrating operational, temporal, and environmental factors; Structured Differential Diagnosis (+18%), prioritising hypotheses with verification steps; and, Decision Optimisation (+22%), balancing conflicting criteria like cost and efficiency. The solution reduces reliance on specific personnel, making expertise accessible for proactive planning and effective decision-making by any worker.

Over the past few years, Large Language Models (LLMs) have revolutionized various fields by leveraging their extensive capabilities. Their ability to generate meaningful and contextually coherent text from minimal inputs, has led to significant advancements in areas such as Natural Language Processing (NLP) and Generative Artificial Intelligence (GenAI). These models have also shown potential in assisting or even automating complex tasks, including code generation. However, as LLMs become increasingly integrated into real-world applications and are made more accessible to the general public, concerns about the security and safety of the content they output have grown significantly. Recent research has shown that these models are vulnerable to adversarial techniques, which can manipulate them into generating harmful or biased outputs. These risks highlight the need for robust evaluation methods to assess the resilience of LLMs against these threats. While some efforts have been done to benchmark LLMs’ robustness against adversarial prompt-based attacks, these approaches are often highly context-specific and lack the comprehensive components required for a robust and systematic evaluation.In this dissertation, we have explored the concepts of LLMs, security, safety, and benchmarking with the goal of establishing a systematic benchmarking methodology for evaluating and comparing the robustness of LLMs built-in security measures when faced with adversarial prompt-based attacks. The proposed methodology is composed of the key components that make up a reliable benchmark, including scenarios, workloads, metrics and attack loads and is designed to ensure its representativeness, usefulness and adaptability to different contexts.To demonstrate the usefulness of the proposed methodology, we have also conducted benchmarking campaign with a varied set of LLMs in the context of vulnerable and also malicious code generation. From this instantiation, we uncovered significant insights into the models’ safety alignments and vulnerabilities. Our findings reveal that while most safety-aligned models effectively refuse to generate malicious code, they readily produce vulnerable code when asked. Furthermore, our analysis of various attack vectors demonstrated that multi-turn conversational attacks and single-turn role-playing scenarios are significantly more effective at bypassing safety measures than template-based prompts. These results underscore the importance of a structured benchmarking methodology and reveal key weaknesses in current LLMs, providing a clear path for future research in developing more robust and secure models.

Accurate text classification and placement remain challenges in U.S. higher education, with traditional automated systems like Accuplacer functioning as “black-box” models with limited assessment transparency. This study evaluates Large Language Models (LLMs) as complementary placement tools by comparing their classification performance against a human-rated gold standard and Accuplacer. A 450-essay corpus was classified using Claude, Gemini, GPT-3.5-turbo, and GPT-4o across four prompting strategies: Zero-shot, Few-shot, Enhanced, and Enhanced+ (definitions with examples). Two classification approaches were tested: (i) a 1-step, 3 class classification task, distinguishing DevEd Level 1, DevEd Level 2, and College-level texts in one single run; and (ii) a 2-step classification task, first separating College vs. Non-College texts before further classifying Non-College texts into DevEd sublevels. The results show that structured prompt refinement improves the precision of LLMs’ classification, with Claude Enhanced + achieving 62.22% precision (1 step) and Gemini Enhanced + reaching 69.33% (2 step), both surpassing Accuplacer (58.22%). Gemini and Claude also demonstrated strong correlation with human ratings, with Claude achieving the highest Pearson scores (ρ = 0.75; 1-step, ρ = 0.73; 2-step) vs. Accuplacer (ρ = 0.67). While LLMs show promise for DevEd placement, their precision remains a work in progress, highlighting the need for further refinement and safeguards to ensure ethical and equitable placement.

Accurate lung cancer diagnosis and staging are critical for personalized treatment and improved patient outcomes. Traditional diagnostic methods, such as manual image examination, are prone to variability and inefficiency. Recent advancements in Deep Learning (DL) have demonstrated potential in automating lung cancer classification and staging, thereby enhancing diagnostic accuracy and efficiency. However, existing solutions often address only isolated aspects of the diagnostic process, such as tumor detection, without offering a unified system for multi-target classification that integrates clinical tools for both clinicians and patients. This study presents a unified framework for lung cancer analysis that combines medical imaging, clinical data, and Large Language Models (LLMs) to support three key tasks: tumor type classification, TNM staging, and automated treatment protocol recommendation. Image-based classification was performed using YOLOv8n, trained on two CT datasets, achieving a maximum mean average precision (mAP50) of 0.418 and an F1 score of 0.44. TNM staging was addressed through a multimodal classifier, combining the ResNet50 model with a multilayer perceptron, which fused imaging and demographic inputs. This approach yielded an average F1 score of 0.389, with the M component showing the strongest performance. For treatment generation, a Retrieval-Augmented Generation (RAG) approach was employed, combining clinical prompts with relevant documents to produce personalized protocols using the Gemini 2.0 Flash LLM. The best-performing configuration achieved a stage match of 0.54 and a BERTScore of 0.83, along with high contextual fidelity across RAGAS metrics (Faithfulness: 0.68, Answer Relevancy: 0.81, Context Precision: 0.99, and Context Recall: 0.93). This framework demonstrates the potential to support both diagnosis and treatment planning within a single integrated system, contributing to more personalized and effective clinical decision-making in oncology.

This thesis explores the practical integration of large pre-trained language models (LLMs) into a real-time web-based recommendation system. While most existing studies emphasize the development and evaluation of new algorithms, there is limited research addressing the implementation challenges faced by non-expert developers seeking to use AI models via APIs. Using the Design Science Research Methodology (DSRM), this project designs, builds, and evaluates a prototype e-commerce platform that delivers personalized product recommendations powered by OpenAI’s GPT-3.5-turbo model. The system collects user behavior data (searches, clicks, favorites) and constructs prompts dynamically to generate real-time product suggestions. The system was tested both technically, through latency and concurrency testing, and with users, who gave feedback on satisfaction, trust, and how accurate the recommendations felt. Results show that the system delivers personalized and timely suggestions with acceptable latency under concurrent use. Overall, the study shows that pre-trained language models are well-suited to support recommendation systems in a simple and effective way, especially for developers who do not specialize in machine learning.

With the rapid evolution of digital technologies, online reviews have become a pivotal source of customer feedback, creating a need for more efficient tools to rapidly analyze customer opinions. This need motivates the exploration of advanced Artificial Intelligence (AI) techniques to enhance customer feedback analysis. This study investigates the efficacy of utilizing Large Language Models (LLMs), for the comprehensive analysis of online reviews. It focuses on evaluating customer feedback through reviews collected from TripAdvisor for hotels in the Algarve region of Portugal. Using the Azure OpenAI API to implement the GPT-3.5-Turbo-16K model for Text Mining tasks such as sentiment analysis and topic modeling. The primary goal of the study is to demonstrate how LLMs can accurately classify sentiments, identify recurring themes, and extract actionable insights from large datasets of unstructured text. The findings highlight LLMs’ ability to provide detailed real-time insights, easily correctly classifying most of the reviews with Positive classifications and highlighting the most frequent pros and cons mentioned, such as “friendly and helpful staff” and “issues with bar service”, providing detailed suggestions to address negative classifications. The research addresses the limitations of LLMs, such as potential biases and the challenge of maintaining accuracy across different languages and cultural contexts. These findings demonstrate the ability to automatically analyze and extract insights from online reviews, enabling more informed strategic decisions, refined marketing strategies, improved product offerings, and enhanced customer satisfaction.

Industrial systems in the context of Industry 4.0 are facing ongoing challenges related to data consistency and interoperability, particularly in Digital Twin implementations. Although the Asset Administration Shell (AAS) provides a standardized foundation for structuring and exchanging industrial data, many systems still rely on custom formats, lim- iting reuse and integration. Creating and updating AAS repositories is time-consuming and requires specialized knowledge, and ensuring interoperability while maintaining flexibility and reusability remains a key challenge. To address these challenges, this dis- sertation proposes the development of a standardized Pose Submodel Template, designed according to the Industrial Digital Twin Association (IDTA) guidelines, to represent robotic poses in a structured and standard-compliant way. In parallel, a multi-agent system pow- ered by Large Language Models (LLMs) was implemented to automate AAS submodel generation, along with a Query–Response subsystem enabling natural language interac- tion with the existing AAS infrastructure, NOVAAS. The results show that combining standardized AAS submodels with generative AI effectively automates submodel creation, reduces human intervention, and facilitates data access, contributing to more scalable, interoperable, and intelligent Industry 4.0 systems.