Economist · Data scientist · Founder
I build systems that put a number on a person — and I decide what that number is allowed to mean.
Seven years of production fraud and risk models in payments, e-commerce and financial services,
where a score decides something about someone and being wrong is expensive in both directions.
A study of 51,858 Brazilian workers on what a screening threshold costs the people it misses.
A formal theory of leader vitality, carried the whole way from the proof to the pip install.
A book about exhaustion, and the company that came out of it — where the instrument answers to
the person who took it and to nobody else.
Formal model, then code, then something a person can actually use. I have built that chain end to end, alone.
Different fields, one question: how much is a number about a person allowed to mean?
- 🎓 B.Sc. Economics USP · M.Sc. Aeronautical Engineering ITA · PhD candidate in AI, ICMC-USP
- 📝 192 citations — Transportation Research Part A (95) · Sustainable Production and Consumption (76) · Scholar · ORCID
- 📚 Author of A Coragem de Parar (2026) — on deciding to stop when no diagnosis arrives to authorize it · PT-BR · English
- 🏆 William Grossman Award for Excellence in Air Transport Research (ITA)
- 🇧🇷 São Paulo, Brazil · building SAPIANS · in Start Digital, Sebrae for Startups
XRESET — a structured self-assessment for people who suspect their work is costing more than it returns. You answer a questionnaire; you get back a written report organized around three questions: how you are, how you experience your relationship with work, and how much room you have to act. It is built to be read once, carefully, and then carried into a conversation with someone who can help. The person answers, the person receives it first, and nothing leaves unless they send it.
Vita — a conversational module that talks only about the report you just received, grounded in your own answers. If the questionnaire's safety items flag acute risk, Vita does not carry on a normal conversation: it opens by saying your safety comes first and offers crisis channels immediately. Safety runs before the conversation, not after it.
Where the instrument stops. A self-report instrument measures dimensions; it does not conclude a case. No diagnosis, no probability of illness, no prognosis, no cause assigned, nothing prescribed — and no individual result ever reaches an employer. Burnout is an occupational phenomenon in ICD-11, not a medical condition, and XRESET does not measure it. Free while the validation studies run.
📐 Dynamic Leadership Vitality Theory (DLVT) — A two-state formal model of how enacted leadership scope, coordination load, and vitality may coevolve. Its central result is the counterintuitive one: narrowing a leader's scope changes what the system can carry and how long adjustment takes, but leaves long-run vitality where it started — relief that lands in the transient, not in the equilibrium. Manuscript complete and checked against the Leadership Quarterly Guide for Authors; not yet submitted. The SSRN preprint is the April 2026 version, posted under an earlier title and a framing the current one abandons.
🧩 Interpretable ML in mental health — Among 51,858 Brazilian workers (PNS 2019), the relationship between weekly hours and depressive-symptom risk is U-shaped with a minimum near 40h, stable across five resamplings. Social support is the strongest protective factor. And a single screening threshold, applied to everyone, buys its average accuracy by missing cases in the groups least likely to be looked at twice. Interpretability cost 0.007 AUC. (ICMC-USP)
📊 interpretable-ml-lectures — Lecture materials for SCC5819 (ICMC-USP), following Molnar. Ceteris paribus, ICE and LIME are published; SHAP is next. Every stated number is printed by the notebook that computes it, and limitations are measured even when they undercut the tidier lesson.
📦 dlvt — The model as installable research code:
simulate vitality–scope dynamics, locate equilibria, classify regimes under a threshold the API
always discloses. Its Scientific boundaries section ends with the only instruction that matters:
use the code to reproduce and challenge the model, do not use it to classify people.
pip install dlvt
I wrote and run the rest privately — the XRESET platform, a deterministic reasoning-memory engine, a grounded-documentation API on Vertex AI, and the multi-agent pipeline that carried DLVT from model to manuscript.
Stack. Python and SQL daily, TypeScript for product. NumPy/SciPy, scikit-learn, XGBoost. PySpark, Databricks and MLflow from the fraud years. Next.js, Supabase, Vertex AI, Terraform. Papers in Typst.
- Fraud, risk & ML advisory — selective, project-based. Seven years of production systems in payments and e-commerce.
- Research collaboration — formal modeling, psychometrics, interpretable ML, or replication of the DLVT results.
- XRESET — take it, or fund access for people who need it. Organizations fund; they never receive individual results.
- Not looking for full-time roles. Running the studio.