Sergio Álvarez Balanyá

I am an Applied Scientist at Amazon working on VLMs and other generative models for image analysis. I did my PhD at the Universidad Autónoma de Madrid under the supervision of Daniel Ramos. I have also taught Machine Learning courses at CUNEF Universidad. My work centers on probabilistic machine learning, including calibration, uncertainty quantification, and Bayesian methods, with applications spanning speech, vision, and language models. I am currently interested in generative models and their application to action prediction and policy learning.

Projects

• Probabilistic Fusion and Calibration of Neural Speaker Diarization Models

  Code for combining outputs from multiple speaker diarization systems using probabilistic fusion and calibration techniques such as Dynamic Logits Fusion and logistic calibration to improve prediction reliability.

• N-Way-Attention

  Explores a generalization of the standard attention mechanism that attends over n-tuples of tokens simultaneously, implementing and evaluating a 3-way attention variant that sums over trios of tokens.

• Adaptive Temperature Scaling

  Implements adaptive temperature scaling for neural network calibration, replacing a single global temperature with a lightweight per-sample function of the logits to better handle varying uncertainty across easy, hard, and near-OOD samples.

• Machine Learning in Numpy

  Collection of Numpy implementations of some traditional ML algorithms. Visual examples are provided in the form of jupyter notebooks.

• Flow Matching from-scratch

  A learning-by-doing exploration of flow matching, progressing from 2D toy problems through MNIST experiments.

All repositories on GitHub →

Courses

• Deep Learning Course

  Hands-on course materials for CUNEF Universidad students, progressing from Python fundamentals to from-scratch neural network implementations and advanced architectures in PyTorch.

• Machine Learning Course

  Course materials for CUNEF Universidad students introducing Python programming, data analysis, and hands-on implementation of core machine learning algorithms.

Selected Publications

• Adaptive Temperature Scaling for Robust Calibration of Deep Neural Networks

  SA Balanya, J Maroñas, D Ramos · Neural Computing and Applications, 2024

• Gaussian Processes for Radiation Dose Prediction in Nuclear Power Plant Reactors

  SA Balanya, D Ramos, P Ramirez-Hereza, DT Toledano et al. · Chemometrics and Intelligent Laboratory Systems, 2022

• Gaussianization of LA-ICP-MS Features to Improve Calibration in Forensic Glass Comparison

  P Ramirez-Hereza, D Ramos, J Maroñas, SA Balanya, J Almirall · Forensic Science International, 2023

• Attribute-Aware Controlled Product Generation with LLMs for E-commerce

  V Negri, VM Gómez, SA Balanya, S Rajaram · arXiv preprint, 2025

• Probabilistic Fusion and Calibration of Neural Speaker Diarization Models

  JI Alvarez-Trejos, SA Balanya, D Ramos, A Lozano-Diez · arXiv preprint, 2025

All publications on Google Scholar →

Writing

• Learnings from Flow Matching, Part 1 Oct 17, 2026

  First part of a series on learnings from implementing flow matching experiments.

All posts → · RSS