Tesis/Trabajos de Grado
URI permanente para esta colección
Incluye documentos como: monografías, reportes, proyectos, prácticas, informes, entre otros; elaborados como requisito de grado para programas de pregrado y posgrado en la Universidad de los Andes.
Navegar
Envíos recientes
Publicación Acceso abierto Herramienta basada en técnicas de Deep Learning y análisis de imágenes satelitales para la estimación del potencial bioenergético de biomasa residual postcosecha de cultivos de café(Universidad de los Andes, 2025-12-11) Castañeda Gutiérrez, Cristhian Camilo; Núñez Castro, Haydemar María; Calderón Romero, Andrés Oswaldo; Giraldo Trujillo, Luis Felipe; Manrique Piramanrique, Ruben FranciscoEste trabajo presenta una metodología para estimar el potencial bioenergético de residuos de café en municipios PDET del departamento del Cesar, Colombia, mediante técnicas de Deep Learning aplicadas a imágenes satelitales Sentinel-2 Level 2A. La detección de cultivos de café representa un desafío técnico considerable frente a otros cultivos como la palma de aceite: las parcelas fragmentadas con promedio de 0.6 hectáreas, las tres modalidades de establecimiento (expuesto, semisombra y sombra) y la topografía accidentada de la Sierra Nevada y la Serranía del Perijá generan una alta heterogeneidad espectral. Para abordar esta complejidad, se implementaron dos arquitecturas de segmentación semántica: DeepLabV3+ con backbone ResNet-101 y SegFormer con backbone MiT-B2. El conjunto de datos se construyó a partir de polígonos del IGAC convertidos en máscaras binarias, aplicando una estrategia de partición Checkerboard Split que garantiza independencia espacial entre los conjuntos de entrenamiento, validación y prueba. Y un desempeño con un IoU = 0.5438. La estimación de producción se realizó mediante un modelo de regresión lineal entrenado con datos históricos de las Evaluaciones Agropecuarias Municipales (2007-2024), obteniendo un coeficiente de determinación R² de 0.89. El cálculo del potencial bioenergético aplica los factores de conversión del Proyecto Nexos (2024), considerando que solo el residuo de hoja tiene disponibilidad efectiva (30%) debido a las prácticas agrícolas locales. Los modelos se integraron en una plataforma web desarrollada con React y FastAPI que permite visualizar los cultivos detectados, estimar la producción y calcular el potencial energético de forma interactiva. Esta prueba de concepto valida la viabilidad técnica del enfoque para su futura extensión a otros cultivos y regiones PDET.Publicación Acceso abierto Identifying Architectural Erosion in Android Apps(Universidad de los Andes, 2025-12-10) Acosta Rojas, Juan Camilo ; Escobar Velasquez, Camilo Andres; Linares Vasquez, Mario ; Garces Pernett, Kelly JohanySoftware engineering involves different efforts with the same purpose: high-quality software development. Among these efforts, we can find the software architecture definition process, which gives preference to the various quality attributes according to the system's needs. However, for different reasons, the solution development can deviate from the original architecture design, which can generate performance problems and, consequently, affect the user experience. This impact on software quality could be represented, at a greater rate, in a mobile environment due to its limited resources. Previous research with this approach has focused on problem study and detection in different areas, like security, connectivity, etc. However, in mobile development, the concept of "architectural erosion" has not been deeply studied. The primary objective of this research project is to identify and locate architectural erosion bugs in mobile applications, utilizing two solution approaches: static analysis code techniques and the application of AI models and NLP fundamentals in commit analysis of Android projects. The results of each methodology will be discussed and extended to solve some issues in Android projects.Publicación Acceso abierto Interoperable Architecture for Digital Identity Delegation for AI Agents with Blockchain Integration(Universidad de los Andes, 2026-01-30) Saavedra Martínez, David Ricardo; Correal Torres, Dario Ernesto; Deng, Juan; Gauthier Umaña, Valérie ElisabethDigital identity infrastructures have evolved powerful mechanisms for authentication and attribute assertion from public key infrastructure and federated identity (OAuth 2.0 / OpenID Connect) to decentralized models based on Decentralized Identifiers (DIDs) and Verifiable Credentials (VCs). However, they still lack a unified, verifiable mechanism for delegation of authority across heterogeneous ecosystems, particularly when autonomous software agents act on behalf of human principals. This thesis proposes a conceptual and architectural framework for verifiable delegation in digital identity systems, designed to operate across centralized, federated, and self-sovereign identity (SSI) environments and to accommodate both human and AI agents. The core contributions are: (i) Delegation Grants (DGs) as first-class authorization artefacts, distinct from VCs, that encode bounded transfers of authority and enforce reduction of scope along delegation chains; (ii) the Canonical Verification Context (CVC), a protocol and format agnostic normalization model that represents any verification request as a single structured object; (iii) a layered architecture that separates trust anchoring, credential/proof validation, delegation and policy evaluation, and protocol routing via a Trust Gateway; and (iv) an explicit treatment of blockchain anchoring as an optional integrity layer, rather than as a structural dependency. The framework is instantiated prototype that accepts heterogeneous inputs (VC-JWT, VC-LD/BBS+, OIDC4VP payloads and DG tokens), normalizes them into CVCs, and executes a deterministic verification pipeline including structural checks, cryptographic validation, delegation chain evaluation, revocation, and optional ledger anchoring. A set of scenario tests and a 1 200 request batch experiment show that verification behaves correctly and deterministically across formats and delegation depths, with latency scaling pproximately linearly with chain length and a bounded overhead when blockchain anchoring is nabled. The research concludes that verifiable, interoperable delegation can be achieved without redefining existing standards, by introducing a normalization layer, an explicit delegation primitive, and a clean separation between verification logic and transport protocols.Publicación Acceso abierto Syndrome Decoding Under Posterior Leakage: Experimental and Theorical Analysis in Post-Quantum Cryptography : Comparative Analysis of Reed-Muller and Random Codes(Universidad de los Andes, 2025-11-28) Florián Quitián, Andrés Felipe; Gauthier Umaña, Valerie Elisabeth; Gauthier Umaña, Valérie; Cardozo Álvarez, Nicolás; De La Cruz, JavierThe transition to the post-quantum era has made the development and evaluation of quantum-resistant cryptographic primitives a critical priority. Digital signatures, in particular, are essential for ensuring authenticity and integrity across modern communication systems. Code-based cryptography is one of the most mature post-quantum families, yet its behavior under realistic leakage conditions remains insufficiently understood. To address this gap, this thesis analyzes leakage-aware decoding strategies and introduces a guided decoding framework that can inform the study of schemes such as Enhanced pqsigRM and, more broadly, Reed–Muller–based constructions exposed to memory remanence leakage. This work examines Cold-Boot Attacks in the context of post-quantum cryptography and investigates how partial, noisy information obtained from side-channel leakage can be incorporated into decoding strategies. Memory decay is modeled through a conditional bitwise posterior model, which assigns per-coordinate reliability estimates based on observed leakage. Building on this model, a novel posterior-guided decoding framework is introduced, featuring a Recursive Posterior-Guided Decoding Algorithm that combines information-set decoding with probabilistic guidance and a genetic enumerator. Experimental results comparing random and Reed–Muller (RM) codes show that RM codes converge faster and more consistently for code lengths n=32 and n=64. While scalability constraints remain at higher dimensions, these findings indicate that the structural properties of RM codes may present unique vulnerabilities to cold-boot leakage compared to random code constructions.Publicación Restringido SciBETO: A Domain-Specific RoBERTa Model Trained on Spanish Scientific Texts(Universidad de los Andes, 2025-11-27) Prieto Avella, Juan Camilo; Manrique Piramanrique, Rubén Francisco; Flórez Fernández, Héctor Arturo; Nuñez Castro, Haydemar María; Facultad de IngenieríaEn este estudio, desarrollamos un codificador especializado en español para textos académicos con el fin de abordar la falta de modelos específicos de dominio en este idioma. Alternativas actuales como BETO y BERTIN se entrenan con corpus de dominio general. Siguiendo el enfoque de SciBERT, que demostró que el entrenamiento con datos específicos de dominio produce mejoras sustanciales en tareas de textos científicos, preentrenamos nuestro codificador con un amplio corpus de publicaciones científicas en español. Posteriormente, evaluamos su rendimiento comparándolo con codificadores de español existentes, utilizando parámetros científicos tanto de dominio general como especializados. Para mejorar la reproducibilidad y promover nuevos avances en el PLN en español, publicamos el corpus de preentrenamiento bajo solicitud y todos los conjuntos de datos de evaluación junto con nuestro modelo en Hugging Face.Publicación Acceso abierto A Formal Verification Framework fo CRDTs in Athena(Universidad de los Andes, 2025-12-15) Ángel Sánchez, Leonardo; Cardozo Álvarez, Nicolás; Varela, Carlos A. ; Escobar Velásquez, Camilo Andrés; Cardozo Álvarez, Nicolás; FLAGlabConflict-Free Replicated Data Types (CRDTs) enable distributed systems to achieve eventual consistency without coordination, making them necessary for applications such as collaborative editors, distributed databases and decentralized systems. Designing CRDTs is complex: subtle choices in merge semantics can lead to unintended divergence, and proving correctness properties such as convergence, commutativity, and idempotence requires rigorous reasoning. While existing work demonstrates that individual CRDTs can be verified using theorem provers, each verification effort typically starts from scratch, requiring substantial proof overhead for common properties. Our proposal offers a structured approach for CRDT verification by developing a modular framework in the Athena proof assistant. The framework facilitates the specification of both the syntax and operational semantics of CRDTs and guides the proof of key properties such as con- vergence, commutativity, and idempotence. By identifying and encapsulating common patterns in CRDT design and verification. It is suited to developers who are exploring or prototyping new CRDT variants and need a precise and checkable way to reason about their correctness. While the framework does not cover all possible CRDT designs. It takes a modest step toward making formal verification more approachable in this domain.Publicación Acceso abierto Generación Semi Automática de Cursos en Seguridad y Salud en el Trabajo mediante Modelos de Lenguaje(Universidad de los Andes, 2025-11-27) Cachique Leguizamón, Sara María; Manrique Piramanrique, Rubén Francisco; Mariño Drews, Olga; Cruz Robayo, Gionvanni Andrés; Facultad de Ingeniería::TICSw: Tecnologías de Información y Construcción de Software; FlagLabEste documento examina el uso de la inteligencia artificial, específicamente los Modelos de Lenguaje de Gran Escala (LLMs), en la generación de contenido educativo, con énfasis en el diseño de cursos personalizados para SST- Seguridad y Salud en el Trabajo. A partir de una revisión de estudios recientes, se identifican avances significativos en la estructuración automática de los cursos, generación temática y acompañamiento formativo con los LLMs. Sin embargo, también se evidencia un vacío recurrente en cuanto a la capacidad de estas herramientas para adaptar los contenidos a contextos específicos, lo cual limita su aplicación en entornos regulados como es la seguridad y salud en el trabajo. Por lo anterior, se desarrolló una herramienta semiautomática que combina las capacidades generativas de modelos de lenguaje (LLMs) con la intervención de asesores humanos, para diseñar contenidos de capacitación personalizados, ajustados a la normativa vigente y a las características específicas de una organización. Esta solución pretende optimizar la planificación de cursos, reducir la carga operativa y administrativa y mejorar la pertinencia pedagógica en entornos de formación de personal para las empresas en los diferentes sectores económicos.Publicación Restringido Evaluating Sentiment Analysis in 19th-Century Illustrated London News: A Multimodal Study(Universidad de los Andes, 2025-12-01) Cohen Solano, Kevin ; Manrique Piramanrique, Rubén Francisco; Facultad de Ingeniería; FLAGLabThis thesis investigates multimodal (image+text) sentiment analysis in a historical setting and compares it with contemporary social media. A curated benchmark of 1,278 engraving–caption pairs from the Illustrated London News (ILN) is introduced, with OCR correction and non-expert sentiment labels, and is evaluated alongside the MVSA-S and MVSA-M Twitter datasets. Five strategies are studied, from a prompt only GPT baseline and a text-only encoder with generated image descriptions to dual-encoder fusion and an A2II-inspired instruction-guided model. On modern data, image+text models consistently outperform text-only baselines, especially for neutral and negative classes, and the most complex architecture yields the strongest results when sufficient high-quality data are available. On ILN, however, gains from visual information are limited and sometimes disappear, and the prompt-only GPT configuration becomes a competitive and stable option under data scarcity. These findings show that multimodal methods can benefit both contemporary and historical sentiment analysis, but that their effectiveness depends critically on corpus design and dataset quality. The work contributes a new ILN benchmark, a systematic comparison of multimodal strategies across domains, and a basis for future extensions to richer historical resources, including full-page newspaper context and 19th-century Latin American press.Publicación Acceso abierto Architectural Foundations for Multistage Task Orchestration in Engineering Educational Systems(Universidad de los Andes, 2025-12-05) González Ruales, Esteban; Escobar Velasquez, Camilo Andres; Lozano Garzón, Carlos Andrés; Padilla Agudelo, Jesse; Facultad de Ingeniería::TICSw: Tecnologías de Información y Construcción de SoftwareThis thesis presents the design of a distributed, fault-tolerant, and extensible system to orchestrate multistage task execution in heterogeneous environments. The system is motivated by the need for an infrastructure capable of processing dynamically routed tasks, initiated by users and subject to variation in outcome and resource requirements, within educational and experimental contexts. A message-driven architecture is proposed in which stateless processing nodes consume tasks from a message broker, retrieve associated artifacts, execute them in isolation, and forward results based on a predefined routing structure. The system is implemented using an actor-based concurrency model, enabling strong fault isolation, structured recovery, and scalable execution through parallel disposable workers. Key architectural principles include the separation of orchestration and execution logic, the asynchronous transmission of messages between components, and the strict isolation between processing tasks. The platform currently supports containerized workloads and provides extension points for integrating new task types by modifying only the executor logic, without altering the orchestration layer.Publicación Acceso abierto Sistema Inteligente basado en ontologías y LLM para la recomendación de recursos de aprendizaje en competencias digitales y de IA(Universidad de los Andes, 2025-12-01) Cruz Manrique, Camilo Andrés; Mariño Drews, Olga; Aguirre Herrera, Sandra Leonor; Manrique, Ruben FranciscoLa acelerada integración de la Inteligencia Artificial (IA) en la sociedad ha redefinido las competencias digitales fundamentales, creando una demanda de formación personalizada que supera las capacidades de los Sistemas de Gestión de Aprendizaje (LMS) tradicionales. Simultáneamente, el uso emergente de Grandes Modelos de Lenguaje (LLM) como asistentes educativos presenta riesgos críticos de integridad factual ("alucinaciones"). Esta tesis propone, implementa y valida una arquitectura de recomendación híbrida tipo Ontology-Grounded Retrieval-Augmented Generation (RAG), diseñada para localizar recursos de aprendizaje abiertos adaptativos y confiables. La metodología abarcó el diseño formal de una ontología de dominio que unifica los marcos DigComp 2.2 y de IA de la UNESCO, la construcción automatizada de un grafo de conocimiento poblado con Recursos Educativos Abiertos (REA) mediante agentes validadores, y el despliegue de un prototipo neuro-simbólico. La evaluación empírica (N=25) evidenció una alta usabilidad del sistema (SUS Score: 80.0) y un nivel de confianza de 4.4/5.0 en la pertinencia de las recomendaciones. Los resultados demuestran que el anclaje ontológico mitiga eficazmente la incertidumbre generativa de los LLM, constituyendo una solución robusta para la curación de contenidos educativos, bajo un compromiso (trade-off) de latencia inherente al razonamiento semántico.Publicación Restringido Neural distinguishers and the challenge of larger-key block ciphers: the case of PRESENT(Universidad de los Andes, 2025-11-28) Martínez Martínez, Isabella; Gauthier Umaña, Valerie Elisabeth; Manrique Piramanrique, Ruben Francisco; Villanueva Polanco, RicardoNeural-differential cryptanalysis has emerged as a powerful technique for evaluating the security of block ciphers. However, existing neural distinguishers face two significant practical limitations that this thesis addresses. First, they are often computationally expensive, and their application to practical key recovery has been confined to ciphers with small round key sizes, leaving their scalability to larger keys—such as the 64-bit round key in the lightweight cipher PRESENT—an open question. Second, these models are notoriously data-hungry, relying on massive training datasets. This thesis first confronts the challenge of computational efficiency and large-key recovery. We introduce a novel entropy-based neural distinguisher for PRESENT. By employing information theory to identify a small, highly relevant subset of ciphertext bits, this method serves as a powerful feature engineering step. The resulting distinguisher uses a network with less than 10% of the parameters of state-of-the-art models while achieving comparable accuracy. We leverage this efficiency to design and demonstrate a practical, iterative key recovery process, constituting the first—to our knowledge—successful, full neural key recovery attack on a 64-bit round key of this nature. Second, this thesis directly investigates the challenge of data efficiency. We explore the use of hybrid quantum-classical models as a direct replacement for the classical distinguisher, trained on the same low-dimensional, entropy-selected bit subsets. We hypothesize that the unique structure of variational quantum circuits offers advantages in a scarce-data regime. Our results demonstrate a significant gain in data efficiency: the hybrid model achieves accuracy approaching 97% of our classical, bit-reduced baseline for PRESENT 6-round and 7-bits while using only 0.2% of the original training data (20,000 examples). In summary, this thesis presents two principal contributions: an entropy-based methodology that yields a computationally efficient distinguisher, enabling the first practical 64-bit neural key recovery attack on PRESENT; and the design of a data-efficient hybrid quantum-classical distinguisher that achieves comparable performance with a fraction of the training data.Publicación Acceso abierto QualiCode AI-assisted qualitative analysis: an agent-based system for iterative coding(Universidad de los Andes, 2025-11-28) Vargas Salamanca, Omar Esteban; Mariño Drews, Olga; Manrique Piramanrique, Rubén Francisco; Núñez Castro, Haydemar Maria; González, Daniel; Facultad de Ingeniería::TICSw: Tecnologías de Información y Construcción de SoftwareQualitative evaluation aims to gain a deep understanding of phenomena, processes, or situations through descriptive data, such as interviews, observations, and documents. Unlike quantitative evaluation, which uses numbers and measurable data as inputs, qualitative evaluation uses de scriptions and observations to answer the “how” and “why” questions. This approach is essential for analyzing complex human and social dynamics. It often relies on manual tasks such as coding and interpreting texts, which can be time consuming and highly repetitive. In addition, the coding process is inherently iterative: re searchers frequently review, refine, or delete codes as they examine the data. This thesis presents the design and implementation of an intelligent support system that leverages large language models (LLMs) and AI-based agents to assist researchers in this iterative process. The system helps generate initial codes, supports refinement based on user feedback, and reveals conceptual relationships between codes by incorporating semantic context. By integrating AI into qualitative research workflows, the proposed approach seeks to improve the efficiency, consistency, and analytical depth of qualitative assessments. The system introduces a flexible and interactive coding workflow that does not seek to replace the researcher but rather to integrate their actions with those of agents and serve as an additional input to the various fields of qualitative researchPublicación Restringido Evaluating language agent architectures for question answering(Universidad de los Andes, 2025-11-28) Rayo Mosquera, Jhon Stewar; Manrique Piramanrique, Rubén Francisco; Cantoral Ceballos, José Antonio; Nuñez Castro, Haydemar MariaQuestion Answering (QA) systems receive a natural language query and are required to provide an answer based on a collection of documents. Recently, Large Language Models (LLMs) have demonstrated strong performance on these tasks. However, certain QA problems remain challenging. Multi-Hop Question Answering (MHQA) tasks, which require gathering evidences from multiple sources, and combining information to produce a correct answer, continue to pose difficulties for LLM-based QA systems. In this work, we design and implement three language agent architectures, DAG Agent, Auto DAG Agent, and Long-lived QA Agent, for solving MHQA tasks based on the Cognitive Language Agents framework. We systematically evaluate the performance of these agents using well-known benchmarks, HotpotQA, 2WikiMultiHopQA, and MuSiQue, and compare them against multiple baseline Retrieval-Augmented Generation (RAG) systems that employ lexical, semantic, and graph-based retrievers, as well as several variants of a simpler ReAct agent. The proposed language agents demonstrate significant higher retrieval than the baseline systems, and competitive QA performance, indicating that they effectively leverage decision-making, reasoning, and reflection capabilities to improve multi-hop retrieval. For instance, using GPT4.1-mini, Auto DAG agent achieves approximately a +5-point improvement in recall on HopotQA and 2WiKi, and a +10-point increase on MuSiQue compared with the baseline agent with GPT-4.1-mini, while exhibiting only a small decline in precision. Similarly, Auto DAG agent achieves competitive performance across the three MHQA benchmarks for Exact Match (EM), ROUGE-1 (R1), and an LLM-based (L1) score. Overall, our results highlight the potential of language agent architectures for MHQA tasks, demonstrating a clear advantage over traditional RAG systems commonly used in QA. Despite these improvements, language agent architectures utilize more resources, posing challenges for real-world scalable applications, and show reduced performance on a more diverse dataset, LoCoMo, suggesting limited generalization to other types of questions.Publicación Acceso abierto Constructing a language for testing Reinforcement Learning programs using NLP techniques(Universidad de los Andes, 2025-07-28) Medina Afanador, Luis Alejandro; Cardozo Álvarez, Nicolás; Dusparic, Ivana; Manrique Piramanrique, Rubén FranciscoProbar el buen funcionamiento de un Agente en Reinforcement Learning es un desafío, esta tesis plantea usar técnicas de NLP para generar casos de prueba en donde la posibilidad de fallos es más probable, interpretando que la interacción entre un agente y el ambiente es un lenguaje o una secuencia según convenga.Publicación Acceso abierto Evaluation of transfer learning for driving in roadside behavior in autonomous robots(Universidad de los Andes, 2025-05-27) Gándara Vega, Lucas; Cardozo Álvarez, Nicolás; Sánchez Puccini, Mario Eduardo; Velasco, Nelson; Velasco, AlexandraWith the development of autonomous driving technology, the need for generalization in the knowledge of autonomous systems has become increasingly important. In reinforcement learning, the time and sample efficiency of the training process is a critical factor on why is not widely used in real-world applications. Knowledge generalization is another constraint to be considered, one simple change in the environment can lead to a complete failure of the system. We analyze the use of transfer learning through the use of a pre-trained model and expert demonstrations on the improvement of the training process of a lane change task in a simulated environment. We use Gazebo and a Turtlebot robot to simulate the lane change paradigm. We established comprehensive benchmarks by training specialized expert agents for both driving paradigms independently. The right-side expert achieved competence in 149 steps (5.82 hours), while the left-side expert reached competence in 3.461 hours. Using these baselines, we quantitatively evaluated transfer learning benefits through three key metrics: 1. Transfer Ratio: The agent utilizing formal transfer learning techniques achieved a transfer ratio of 0.9401, indicating substantial knowledge preservation between paradigms. This was superior to the agent learning both paradigms simultaneously (1.1455). 2. Mastery: Our transfer learning agent maintained high performance levels comparable to independently trained expert agents, demonstrating that knowledge transfer did not compromise behavioral quality. 3. Pedagogy: Transfer learning reduced training steps by approximately 53% compared to training from scratch (70 versus 149 steps), representing significant computational resource savings. Most notably, the agent implementing gradient transfer equations achieved competence in just 1.929 hours, representing a 44.3% reduction in training time compared to learning from scratch (3.461 hours). Interestingly, the agent that learned from the right expert without formal transfer learning techniques achieved an even higher maximum performance level, suggesting potential benefits in exploring hybrid approaches. Our research demonstrates that transfer learning offers a powerful approach for enhancing autonomous driving systems’ adaptability and efficiency, bringing us closer to developing vehicles capable of seamlessly adapting to diverse driving conditions while significantly reducing computational resources required for training. Our research demonstrates that transfer learning offers a powerful approach for enhancing autonomous driving systems’ adaptability and efficiency, bringing us closer to developing vehicles capable of seamlessly adapting to diverse driving conditions while significantly reducing computational resources required for training.Publicación Acceso abierto Enhancing gender violence legal guidance with LLMS and retrieval systems(Universidad de los Andes, 2025-06-20) Martínez Carrión, Santiago; Manrique Piramanrique, Rubén Francisco; Molano Giraldo, María Fernanda; Rueda Rodriguez, Sandra JulietaThis study investigates recent developments in AI applications for the legal domain, emphasizing the capabilities of LLMs, legal chatbots, and retrieval-augmented generation (RAG) to bridge the gap between legal systems and the general public. To support this research, a synthetic dataset of legal conversations was developed through simulated interactions between lawyers and users, curated in collaboration with legal experts. Furthermore, we propose an agent-based architecture designed to: (1) ensure comprehensive extraction of case details from user interactions, (2) systematically construct legal cases from this information, and (3) integrate contextually relevant legal frameworks via RAG to generate useful guidance for users.Publicación Acceso abierto Explorando el uso de la Realidad Virtual y el análisis de datos para el bienestar emocional en un entorno universitario(Universidad de los Andes, 2025-05-28) Mora Valbuena, Julián Camilo; Figueroa Forero, Pablo Alejandro; Mariño Drews, Olga; Gutiérrez Vela, Francisco Luis; Facultad de Ingeniería::Imagine: Computación Visual, I+D+IEsta investigación evalúa el potencial de las tecnologías de Realidad Virtual (RV) como herramientas para promover el bienestar emocional en entornos universitarios, analizando su influencia en la gestión emocional y el estado energético de los participantes. El trabajo se estructuró en tres fases: diagnóstico e identificación de alternativas, desarrollo y evaluación experimental. En la fase de diagnóstico, se estudió la literatura especializada y se estableció contacto con el área de Seguridad y Salud en el Trabajo (SST) para identificar áreas de acción efectivas. Durante la fase de desarrollo, se generaron dos vías de intervención: una basada en software comercial y otra en desarrollos propios que permitieran mayor control en el análisis de datos. Se crearon dos aplicaciones originales para Meta Quest 3 con enfoques complementarios: Night Forest, orientada a la introspección y la regulación emocional mediante experiencias contemplativas, y Mood Room, centrada en la canalización energética a través de minijuegos que promueven el movimiento del usuario. Ambas aplicaciones integran un sistema de telemetría que registra patrones de atención, tiempo de uso, desempeño y otros indicadores relevantes, preservando la privacidad de los usuarios.Publicación Acceso abierto A framework for the analysis of quantum safe cryptography(Universidad de los Andes, 2025-06-20) Ruiz Gómez, Julio César; Cardozo Álvarez, Nicolás; Gauthier Umaña, Valerie Elisabeth; Perez Bernal, Juan FernandoDiseño y desarrollo de un framework para evaluar la eficiencia espacial y temporal de algoritmos de criptografía post-cuántica.Publicación Restringido A semi-automated approach to Colombian political ontologies through LLM agents and journalist analysis(Universidad de los Andes, 2025-06-25) Rodríguez Fonseca, José David; Manrique Piramanrique, Rubén Francisco; Rueda Rodriguez, Sandra Julieta; Uriza Antorveza, Pablo Andrés; FLAGThe combination of Knowledge Graphs (KGs) with Large Language Models (LLMs) is known as a paradigm shift in information retrieval and natural language understanding systems. This research proposes a novel methodology for semi-automatic generation and validation of KGs through the orchestration of LLMs and complementary frameworks, with a specific focus on Spanish language applications. Given the absence of high level annotated Spanish datasets, this investigation leverages advanced techniques, including LLM agent graphs, to facilitate ontology learning (OL) processes and automate document generation within journalistic contexts. The methodological framework encompasses systematic identification of data sources, implementation of extract, transform, and load (ETL) processes, database architecture design, selection and optimization of baseline models, and the development of an agent interaction protocol. The study explores a multi-model pipeline incorporating Named Entity Recognition (NER), Relation Extraction (RE), and Event Extraction (EE), subsequently integrated into Resource Description Framework Schema (RDFs) or Ontology Web Language (OWL) format. Through a comparative evaluation against contemporary state-of-the-art approaches in Knowledge Graph generation and ontology learning systems, this research advances our understanding of scalable KG solutions.Publicación Acceso abierto Resilient routes: leveraging deep learning to identify critical transportation links for humanitarian response(Universidad de los Andes, 2025-06-03) Aguirre Salamanca, Daniel; Herrera Suescún, Lizeth Andrea; Cardozo Álvarez, Nicolás; Macea, Luis; Manrique Piramanrique, Rubén FranciscoThis research presents a novel application of graph neural networks (GNN) for identifying critical links in transportation networks during disaster response operations. We simulated diverse transportation networks with varying topologies, weight patterns, and source-terminal configurations to construct a comprehensive training dataset. For each network, we calculated critical link scores using three established vulnerability indexes—Cantillo’s index, global efficiency, and number of independent paths. Traditional approaches are computationally intensive, relying heavily on network size, supply/demand nodes, and disruption scenarios, making them impractical for large networks. Our GNN approach addresses this limitation by learning vulnerability patterns directly from network structures, eliminating the need for exhaustive scenario evaluation during inference. To validate the proposed approach, we applied it to both simulated and real-world transportation networks with diverse topological characteristics. We assess performance by sequentially removing links according to their criticality rankings and measuring the Global Importance Index (GII), which quantifies unsatisfied demand and importance post-disruption. This sequential disruption analysis allows direct comparison between GNN predictions and traditional vulnerability indexes. Comparative timing analyses demonstrate that our model identifies critical links with comparable accuracy but at a fraction of the computational cost, even under varying disruption scenarios. This work offers a scalable methodology for conducting vulnerability assessments on previously unmanageable network scales, advancing the planning of humanitarian aid distribution during natural disasters.