Surrogate models for finding optimal inputs
Gaussian processes, Bayesian optimization and Expected Improvement for finding strong configurations under a limited evaluation budget.
José Arbelaez
Data scientist & developer
Data, models and systems worth building.
Projects
RoadGuard
Reliable road damage detection: cross-country transfer, confidence calibration, abstention policy and explainable maintenance prioritization.
BGG Review Intelligence
From BoardGameGeek reviews to structured opinion knowledge: linguistic preprocessing, sentiment classification and aspect-based analysis.
Image generation with diffusion models
Implementation of the SDE framework by Song et al.: denoising score matching training, samplers, conditional generation and imputation on CIFAR-10.
Built and shipped
Tools and experiments that are up and running, with the code in the open.
Backtesting and portfolio analysis
2025
The toolkit my team used to win the UAM Investors League: first among 240+ participants, with the competition's highest Sharpe ratio. Sharpe and Sortino, drawdown, VaR, efficient frontiers and Monte Carlo scenarios.
Reinforcement Learning Lab
Jan - Sep 2024
From tabular Q-learning and SARSA on FrozenLake, with Lorena Mohanu, to DQN and PPO style components in PyTorch. The focus is stability, sample efficiency and what actually changes an agent's behaviour.
Endy and Multiplexor
Runs coding agents (Codex, Claude, Gemini, OpenCode) in managed tmux sessions, tracks parallel tasks and hands work over when one fails or hits its quota. Published on npm, peaking at 5,000 downloads. Multiplexor routes between local providers by priority and quota.
— total views — online