Félix Suárez

The explosive growth of wireless devices has intensified the demand for efficient spectrum management solutions. This work addresses this challenge through Neural Architecture Search (NAS) applied to WiFi spectrum sensing, using 320,000 real IQ signals from the DeepSense dataset. Our approach discovered an ultra-efficient architecture with just 13,412 parameters that achieves 98.4% classification accuracy—a 92.7% reduction in model complexity compared to a baseline with 184,000 parameters and 96.6% accuracy. When deployed on an iPhone 17 Pro Max, the model performs remarkably well with 0.12 ms inference latency and 98.0% accuracy, processing over 8,500 signals per second. This represents a 121-fold speedup compared to training conditions and outperforms even specialized FPGA implementations.

This study presents a comprehensive investigation into optimizing a SQL Server in a business production environment that was experiencing severe performance problems and frequent system freezing. We implemented a systematic methodology consisting of three key phases: thorough diagnosis, strategic optimization, and longitudinal monitoring. Our findings demonstrate statistically significant improvements across critical performance indicators, including a 26% reduction in CPU utilization, a 4.3% increase in available memory, stabilization of the Page Life Expectancy indicator at optimal values exceeding 300 seconds, and complete elimination of system freezing episodes. This research provides a transferable methodological framework for SQL Server optimization across diverse business environments while contributing valuable insights to the efficient management of relational databases in high-availability contexts.

This paper presents a Neural Architecture Search (NAS) approach for automatic classification of telecommunication signals using Convolutional Neural Networks (CNNs). Our optimized CNN architecture achieved 96.29% accuracy using Neural Architecture Search with only 91,367 trainable parameters, demonstrating improved computational efficiency compared to manually designed architectures. We explored 63 million possible architectural configurations using Bayesian optimization to systematically identify optimal network designs for spectrogrambased signal classification. The search process evaluated 20 different trials over approximately 1 hour and 21 minutes, ultimately discovering an architecture that balances accuracy and computational efficiency. This work demonstrates how automated architecture search can effectively optimize telecommunication signal processing applications for spectrum monitoring, electronic defense, and cognitive radio systems.

This paper presents a system for automatic classification of telecommunication signals using signal processing, multi-domain features fusion, and machine learning techniques. Our system achieves a 97.72% classification accuracy across a wide range of SNR values (-20 dB to 18 dB) using an over-the-air radio-frequency (RF) signals dataset, while maintaining a relatively low complexity (167k learnable parameters). We employ a comprehensive feature extraction methodology that combines time-frequency representations, wavelet transform coefficients, and frequency domain statistics which are processed through a multi-layer architecture. This work demonstrates a systematic approach to signal classification that balances accuracy, computational efficiency, and generalization capability, with potential applications in spectrum monitoring, electronic defense, and cognitive radio systems.

An emotion analysis service, based on Transfer Learning, was built using BETO, and applied to Facebook posts written in Spanish. The system was tested using comments related to the anniversary of major media outlets in Costa Rica, the comments on these posts provide insight into the public opinion towards the mentioned media outlet. After collecting the comments and processing them into the emotion analyzer, the most common feelings for each media where: Amelia Rueda (joy), La Nación (anger and sadness) and Teletica.com (joy and anger).

This paper gives an overview of a robotic architecture meant for skillful manipulation. This design is meant to close the gap between the high-level layer (reasoning and planing layer) and the object model system (physical control layer). This architecture proposes an interface layer that allows, in a meaningful way, to connect atomic tasks with controller inputs. In this paper, we discuss how specific complex tasks can be resolved by this system; we analyze the affordance unit design and, we overview the future challenges in the implemenation of the whole system.

In this paper, we propose a high level layer able to translate motion commands in natural spanish language to a formal intermediate representation called Robot Control Language (RCL). The layer was built by using the seq2seq TensorFlow library, with a single forward LSTM for encoder and decoder respectively. We were able to achieve a 4.3e-08 loss employing a manually generated corpus in Spanish.

In this paper we analyze the vocabulary of an action description language to be used, together with a low level object controller, by a humanoid robot in two specific scenarios: a kitchen and a collaborative workshop table. A methodology to acquire this knowledge is presented and then it is applied to these two scenarios. The collected data allows to establish a priority of which actions needs to be implemented with robot controllers, which objects need to be manipulated and how much variation they exhibit between different scenarios or different complex tasks.