Juan Da Silva

Research and development of enhanced AI models for autonomous Fault Detection, Isolation, and Recovery (FDIR) in spacecraft digital payload processors.

Developing novel hardware-aware neural rendering methods, CUDA kernel optimizations, and frequency-domain representations for text-to-3D generation in the group of Prof. Daniel Cremers under the supervision of Dr. Héctor Andrade Loarca.

Conducted research with Prof. Dr. Vincent Heuveline on deep learning for neonatal disease screening and explainable AI (XAI) for clinical interpretability.

Led teaching assistant sessions for linear algebra and numerical linear algebra courses under Prof. Julio César Daza Rodríguez.

Core coursework: machine learning, 3D & multiple-view computer vision, visual recognition, natural language processing, graph neural networks, sequential models, and deep generative models.

Core coursework: numerical linear algebra, optimization for machine learning, probability theory, and mathematical statistics.

Core coursework: fundamentals of machine learning, advanced machine learning, and deep computer vision.

Investigated the training dynamics of infinite-width neural networks. Validated the theoretical bounds of Neural Tangent Kernel (NTK) linearization and Gaussian process convergence in practical, finite-width settings.

Developed a simulation-based Bayesian inference framework (BayesFlow) combining CNNs, LSTMs, and Invertible Neural Networks (INNs) to estimate parameters of complex epidemiological models from partial time-series data.

Designed a U-Net-based crowd-counting architecture featuring a multi-scale inception decoder, mitigating scale variation and perspective distortion in density map estimation.

Architecting and developing an institutional-grade equity research terminal that unifies automated financial data pipelines, quantitative valuation and risk modeling, and interactive visual analytics. Built as a high-performance, containerized web platform featuring dynamic cross-sectional benchmarking, fundamental statement normalization, and sub-second query execution.