Researcher in machine learning and operations research · Applied AI engineer

My research asks when a standard result can be trusted: whether an evaluation protocol changes the scores it reports, whether a clustering survives a small change of its parameter, and whether a simple lower bound already proves a schedule optimal. Each study ships with its code, data and the checks needed to reproduce it. I hold a Master's degree in Computer Science and Computational Mathematics from the University of São Paulo.

• Why Stratified Few-Label Evaluation Changes the Scores of Tabular Foundation Models: An Exact Bayesian Account. Preprint, doi:10.5281/zenodo.23967812.
• stratified-few-label-protocols: code, analysis plans, stored results and regeneration scripts for a study of how stratified few-label protocols can overstate macro-averaged metrics.

• Why HDBSCAN*'s flat clustering changes with minPts: shrinking clusters, a one-run predictor and a certificate. Preprint, doi:10.5281/zenodo.24039671.
• minpts-check: a Python package that checks, from a single HDBSCAN* run, which clusters are at risk when min_samples changes. It works with sklearn.cluster.HDBSCAN and the hdbscan package.

• Lower bounds explain why the distributed assembly permutation flowshop benchmark is easy. Data and code, including the schedules and their lower bounds: doi:10.5281/zenodo.24038696.
• Union-of-predecessors bounds and a forward–backward local search for multi-door cross-dock truck scheduling. Code and data: doi:10.5281/zenodo.23933867.

Fixes and documentation I contributed upstream while doing this research.

• coin-or/SYMPHONY #248: fixed four bugs that crashed the VRP application on master or made it return wrong optima when branching on cuts (a NULL MIP-info dereference, a build-flag mismatch, user cut candidates overwritten by a built-in rule, and a ranged-row convention).
• mhahsler/dbscan #101: documented how to set HDBSCAN's minimum cluster size separately from minPts.
• TutteInstitute/fast_hdbscan #94: clarified that min_samples does not count the point itself.

Bug reports with reproductions: NVIDIA/cuml #8879 (HDBSCAN with nn_descent), fast_hdbscan #93 (cluster_selection_epsilon returning the root), scikit-learn-contrib/hdbscan #757 (alpha ignored by the Borůvka algorithms).

• lgc: an implementation of Learning with Local and Global Consistency, the graph-based semi-supervised algorithm of Zhou et al. (2004).
• ecommerce-data-analyst-agent: conversational data analysis for e-commerce with CrewAI.