Indexación Académica

Publicaciones, Preprints & Datasets

Artículos revisados por pares en conferencias y revistas IEEE/ACM en las áreas de Inteligencia Artificial Multimodal, Neurocomputación y Accesibilidad Sensorial.

IEEE Transactions on Neural Networks and Learning Systems (TNNLS) · 2026

star Destacado

Dynamic Cognitive Load Estimation via Multi-Modal Eye Tracking and Electroencephalography Transformers

Joan Gregorio Pérez * iD

Dr. Elena Rostova iD

Sarah Lin, Ph.D. iD

Understanding real-time cognitive workload is critical for designing adaptive educational interfaces and high-stakes decision-support systems. We propose CogFormer, a novel multimodal cross-attention transformer architecture that synchronizes non-invasive 64-channel EEG streams with high-frequency eye gaze fixations and pupillometry. Across a cohort of 140 participants performing complex algorithmic tasks, CogFormer achieved a 94.2% F1-score in detecting cognitive overload states with a sub-18ms inference latency on consumer-grade hardware.

#Multimodal EEG & Vision #Transformer Architectures

arrow_forward Ver Detalles

ACM Conference on Human Factors in Computing Systems (CHI 2025) · 2025

star Destacado

AURA: Adaptive Sensory-Inclusive User Interfaces for Neurodivergent Children

Joan Gregorio Pérez * iD

Prof. Carlos Mendoza iD

Digital learning platforms often impose severe sensory overstimulation on children with Level-1 Autism Spectrum Disorder (ASD). This work introduces the AURA architectural framework, a zero-shadow, low-contrast, fixed-grid sensory regulation interface that dynamically modulates acoustic, spatial, and visual stimuli based on real-time biometric indicators. Clinical trials over 6 months across 8 inclusive classrooms demonstrated a 41% decrease in sensory meltdown frequency.

#Neurodivergent HCI #Sensory-Inclusive UI

arrow_forward Ver Detalles

Neural Information Processing Systems (NeurIPS 2024 Workshop on Edge AI) · 2024

Sub-Bit Precision Quantization and Weight-Activation Slicing for Edge Transformer Inference

Joan Gregorio Pérez * iD

Marco Aurelio Silva iD

Deploying large transformer models on embedded hardware is severely bottlenecked by memory bandwidth and thermal throttling. We introduce WAS-Quant (Weight-Activation Slicing Quantization), an asymmetric 3.2-bit quantization algorithm that preserves perplexity within 0.12 of FP16 baselines while reducing KV-cache footprint by 68%.

#Edge Quantization #LLM Inference Optimization

arrow_forward Ver Detalles

ACM Technical Symposium on Computer Science Education (SIGCSE 2024) · 2024

Pedagogical Socratic Feedback Generation in Computer Science Education using Fine-Tuned CodeLLMs

Joan Gregorio Pérez * iD

Prof. Carlos Mendoza iD

Automated grading tools frequently reveal solution code prematurely, robbing students of critical problem-solving epiphanies. We present Socrates-Code, a fine-tuned 14B model that generates tiered heuristic questions instead of answers. Longitudinal evaluations across 3 semesters of Algorithms courses demonstrated a 28% increase in retention and debugging proficiency.

#AI in Computer Science Ed

arrow_forward Ver Detalles