This thesis develops a computer-vision system for measuring self-directed head-strike movements in children with autism. Using Azure Kinect recordings, it compares depth-based body tracking, YOLO11-pose, and MediaPipe, incorporating Kalman filtering to estimate hand trajectories, velocity, and deceleration. Results identify depth-based tracking as the most reliable approach, providing a foundation for future impact-force estimation

This thesis develops a hybrid recommendation engine for assigning SAP access roles at CMPC, addressing manual processing and operational bottlenecks. Using a static dataset from Brazilian branches, it combines kernel principal component analysis and collaborative filtering with Gradient Boosting classifiers. The second-stage filter reduces false-positive recommendations, producing a more manageable set of candidates for access-management decisions.

This thesis develops the frontend of Shefu, a mobile application for gamified culinary learning, addressing limited guidance in home cooking. Built with Flutter and a modular architecture, it combines step-by-step instruction with levels, badges, and rankings. User testing reports strong acceptance and ease of navigation, supporting the feasibility of an intuitive interface for independent culinary learning.

This thesis develops a lightweight licence-plate recognition module for automated vehicle-access control. It combines IWPODNet plate detection, character segmentation, and MobileNet recognition, demonstrating extraction and processing of multiple plates from a single image. The implementation supports future integration into a surveillance platform, while validation on Chilean plates and testing with the complete system remain pending. 

This thesis develops a Vision Transformer architecture for detecting and characterising lesions in mammography images, supporting breast-cancer assessment. It adapts pretrained models through transfer learning and fine-tuning, creating three independently trained classifiers. Together, these identify lesion presence, lesion type, and the corresponding BI-RADS category, contributing a structured approach to automated mammographic image analysis.

This thesis develops PedagogIA’s prompt-engineering module and serverless backend, supporting teachers in producing personalised educational materials. It combines iterative human feedback and structured few-shot templates with AWS Lambda, API Gateway, PostgreSQL, and DynamoDB. Nine specialised educational services demonstrate technical feasibility, with concurrency tests reporting average content-generation times of 3.5 seconds and an architecture supporting interchangeable language-model providers.

This thesis develops a digital platform for querying healthcare data through large language models, reducing the need for specialised database knowledge. It translates natural-language questions into SQL and integrates Amazon Bedrock with session management, access controls, conversation records, and satisfaction surveys. The functional prototype also generates charts and provides tools for analysing user feedback and model performance.

This thesis investigates automated mammography classification using convolutional neural networks, comparing EfficientNetV2, ResNet50, and Xception on CBIS-DDSM data. It combines lesion-patch pretraining, transfer learning, and ensemble methods. Results show improved performance from patch pretraining for some architectures, while distinguishing benign from malignant lesions remains challenging and the models require further improvement before clinical deployment. 

This thesis evaluates unsupervised anomaly detection for identifying cybersecurity risks in corporate server networks. It combines Isolation Forest with defensive risk factors to distinguish ordinary findings from potentially higher-risk vulnerabilities and assets. Experiments show favourable results for individual vulnerability detections, while asset-level risk assessment performs less accurately, highlighting limitations in aggregating network-security risk.

This thesis develops a predictive model for identifying patients who may require oncological genetic counselling, in collaboration with FALP. It combines clinical-data preprocessing with BioBERT to analyse structured and unstructured information. The study reports 75.59% precision and integrates Power BI visualisations, supporting interpretation of predictions and more informed decisions about clinical referrals.

This thesis develops a functional implementation of Proximity 2.0 to support differential diagnosis using chest CT examinations. Guided by radiologists’ needs, it integrates medical-image retrieval, clinical information, and a data-preprocessing pipeline. The resulting minimum viable product retrieves visually similar diagnosed cases, supporting clinical comparison while identifying retrieval and processing speed as areas for further optimisation

This thesis investigates attention-based models for detecting anomalies in three-dimensional chest CT examinations. It evaluates Vision Transformer features and transfer learning from natural images, including strategies for combining information across slices. Results demonstrate useful transfer to two-dimensional medical images, but unsatisfactory three-dimensional classification, highlighting the need for segmentation, improved aggregation, and architectures that preserve spatial information.

This thesis develops a natural-language processing model for detecting genetic-risk indicators in breast- and colon-cancer clinical records. Using BioBERT, it identifies red flags that may justify referral for oncological genetic counselling. Interactive Power BI dashboards present predictions and performance metrics, contributing a tool for exploring clinical information and supporting more informed patient-referral decisions.

This thesis investigates machine-learning identification of enzymes capable of degrading aromatic pollutants from genomic sequences. It compares SVM, nearest-neighbour, and Random Forest classifiers across six datasets with different class balances and sequence lengths. Although initial evaluation metrics are promising, testing on complete bacterial genomes produces many false positives, highlighting dataset representativeness and robust validation as central challenges.

This thesis develops a product recommendation system using natural-language processing to personalise users’ experiences. It combines keyword extraction, BERT-derived embeddings, cosine-distance nearest-neighbour retrieval, and TF-IDF reranking. Evaluation reports nDCG of 76% and MRR of 79%, demonstrating the potential of semantic representations and ranking techniques to generate relevant recommendations from user and product information

This thesis evaluates network infrastructure for the CONDOR cosmic-ray and gamma-ray observatory, addressing uncertainty about data-transmission capacity and real-time operation. It uses network simulations, traffic analysis, and transfer-time measurements to assess performance and equipment requirements. The study identifies capacity and latency constraints, providing a basis for infrastructure improvements needed to support detector communication and data processing. 

This thesis develops a pipeline for analysing white-dwarf spectra through neural-network latent spaces. It combines principal component analysis with clustering methods including DBSCAN, K-means, and spectral clustering. The modular pipeline generates visual comparisons with known stellar classes, preserves at least 95% of variance, and processes the tested configurations in under 23 seconds. 

This thesis investigates late fusion of similarity distances to compare complete CT examinations, extending Proximity’s retrieval of individual slices. Using ten chest CT studies, it implements and compares several fusion methods in Python and MATLAB. It contributes a practical framework for sampling, combining, and evaluating retrieval rankings, supporting future development of examination-level medical-image search.

This thesis develops a machine-learning architecture for characterising enzymes associated with aromatic-hydrocarbon degradation. It searches amino-acid sequences for patterns that indicate possible enzymatic functions, supporting the identification of microorganisms useful in industry and research. The work covers data collection and preparation, model training, accuracy evaluation, and analysis of the resulting architecture’s limitations and potential improvements.

This thesis develops a web and mobile platform for coordinating security in the V21 Innovation District. Guided by field visits, stakeholder interviews, and multidisciplinary design, it supports emergency reporting, communication, maps, and event history. Prototype testing identifies usability improvements, while limited connectivity, equipment, and sensor data remain barriers to broader smart-city integration.

This thesis develops a scalable web platform for meal planning at Natgio Kindergarten and Nursery, replacing spreadsheet-based procedures. Built with Quasar, Express.js, and MySQL, it manages ingredients, recipes, and daily menus by preschool level. The deployed solution supports distinct user roles, multiple branches, ingredient-stock reporting, and PDF reports, centralising the institution’s meal-planning workflow.

This thesis investigates automated classification of electrical-grid operating states using data from Chilquinta’s Miraflores substation. It compares CNN, LSTM, and Random Forest models for detecting failures in transformer time series. Results reveal that inconsistent and temporally misaligned failure records significantly limit performance, highlighting data quality as the main challenge for automation. 

This thesis develops a mobile application for corporate emergency management, improving VIPer’s existing mass-notification system. It introduces bidirectional interaction through scenario activation, an information module, and notification history. The solution integrates Flutter, Laravel, cloud services, and push notifications, enabling authorized users to trigger emergency scenarios directly from mobile devices.

This thesis develops a machine-learning support tool for preventive maintenance of Chilquinta’s electrical insulators. Using inspection data, it combines Z-Score and Isolation Forest anomaly detection to identify abnormal measurements and improve data labelling. The resulting system improves data quality, inspection records, and maintenance analysis while requiring minimal changes to existing procedures.

This thesis develops an improved version of Tidys, an educational web application designed to strengthen student–teacher interaction and support active-learning methodologies. It improves usability and scalability while collecting classroom interaction data, enabling researchers and instructors to analyse attendance, attention, participation, and student performance and evaluate teaching methodologies.

This thesis develops the microservice communication architecture for a configurable targeted-advertising recommendation system. It compares ZeroMQ, Apache Kafka, ActiveMQ, and RabbitMQ, selecting and integrating a messaging solution into a functional prototype. Stress, scalability, and failure-recovery tests evaluate whether the resulting architecture provides robust and scalable communication between recommendation-system components.

This thesis proposes a Fourier synthesis method from non-uniform samples using Gaussian Process Regression (GPR), motivated by radio-interferometric imaging. It reconstructs images directly from irregular Fourier-space measurements, avoiding conventional gridding. Experiments show that, with suitable hyperparameters, GPR can outperform FFT-based reconstruction in image error and exhibits strong robustness

This thesis develops a Markovian recommendation engine for digital debt collection. It models collection strategies as a finite-horizon MDP, using machine-learning classifiers to estimate transition probabilities from customer data. Value iteration then recommends short-term collection actions. Experiments and simulations demonstrate the approach’s feasibility while highlighting sensitivity to classifier performance and data imbalance.

This thesis develops a predictive maintenance approach for failure forecasting using supervised deep learning. It estimates the Remaining Useful Life (RUL) of machinery from multivariate sensor data, using NASA’s turbofan degradation dataset. Modified deep-learning architectures are evaluated, achieving improved prediction performance and supporting maintenance decisions based on equipment operating conditions. 

This thesis develops a method for automatically detecting and classifying astronomical sources in large images. It combines image-processing and machine-learning techniques to identify relevant structures and characterize their properties. Experiments on astronomical datasets evaluate detection accuracy and computational performance, supporting automated analysis as data volumes from modern observatories continue to grow.

This thesis investigates efficient indexing and searching of Virtual Observatory resources using succinct data structures. It compresses astronomical XML documents while allowing direct in-memory queries, exploring trade-offs between storage and

This thesis develops an automated process for detecting and classifying long-period variable stars using machine learning. It first identifies long-term variability from astronomical time series, then characterizes stars using physical properties and dimensionality reduction. Clustering with K-means, based on features related to the H-R diagram, identifies six distinct groups.

This thesis proposes automating the ingestion and publication of astronomers’ research data in the Chilean Virtual Observatory (ChiVO). It defines user roles, accepted astronomical data and metadata, and a protocol that automatically generates DaCHS configurations following international IVOA standards. A functional prototype demonstrates feasibility and supports development-cost estimation. 

This thesis evaluates alternative communication middleware for ALMA Common Software (ACS), whose CORBA-based architecture faced challenges for newer observatories such as CTA. It proposes a systematic selection methodology based on PECA, combining multicriteria analysis, candidate filtering, latency benchmarks, and software-coupling analysis to guide future ACS modernization decisions.

This thesis proposes a method for identifying hierarchical, overlapping structures in astronomical images and spectroscopic cubes. Data are represented continuously using Gaussian mixtures, then decomposed through Gaussian Mixture Reduction into a hierarchical tree. Experiments with ALMA and synthetic data demonstrate effective source identification, noise reduction, and representation of blended emissions.

This thesis explores visualization of astronomical spectroscopic data cubes, particularly ALMA FITS files. It compares Matplotlib, PyQtGraph, and Mayavi, selecting Mayavi for efficient 3D rendering. A Python/Qt application combines volumetric visualization, spectra, metadata, and interactive controls. Astronomer validation confirmed its usefulness while identifying improvements for larger datasets and interactivity.