I spent 6 years doing credit risk analysis at a factoring company. Manual work — prospecting clients, evaluating receivables, assessing default risk from gut feeling and spreadsheets. I got bored. I knew the job could be automated. So in 2019 I co-founded Duplify with two friends to rebuild the entire factoring system from scratch.
That's when I started programming. 2020 — first line of Python, first ML model. I taught myself statistics, linear algebra, and machine learning to build a credit scoring system on CNPJ and receivables data using XGBoost, Random Forest, survival analysis (Kaplan-Meier, Cox proportional hazard), and rule-based scoring (ACIUM rules engine). The system worked. Getting clients was the hard part.
In 2022, while studying time series and ML with a post-doctoral researcher at Unicamp, I showed him the credit risk project. He liked it enough to offer me a Master's position. I didn't take it — bureaucracy got in the way — but the recognition confirmed I was on the right track.
In 2023, I co-founded bee6 with four partners. We shipped real estate ML, agricultural AI, restaurant analytics, and document intelligence systems for clients across Brazil and the US. Most experiments failed. What survived is what you'll find in the repos below.
I am also the father of an autistic child — which is why I built una-edu: an AI that generates fully personalized educational materials for autistic students, anchored in their hyperfocus themes.
| una-edu |
Multi-agent pipeline for autistic students: Researcher → Neuro-educational Adapter → Art Director → HTML Renderer. Print-ready A4 materials personalized by grade, hyperfocus theme, and ASD profile. Deployed on Cloud Run. |
| urban-space |
Real estate ML — KNN property matching engine, multilevel pricing model integrating IPTU + RAIS + 10 macro indicators, PCA terrain scouting over 11M+ São Paulo parcels |
| solo-inteligente |
Agricultural AI — automates the "perdigueiro" process for São Paulo's 2023 densification plan; FAISS vector store over INCRA/CAR/SIGEF documents |
| bquant |
Quantitative research toolkit — ADF/KPSS stationarity, ARIMA, ARCH-LM volatility, FipeZAP/Selic/IPCA macro pipelines, B3/CVM data ingestion, KD-Tree spatial joins |
| risco-credito-ml 🔒 |
Credit risk engine for Cicor Factoring — ACIUM rule-based scoring, Kaplan-Meier survival curves, Cox proportional hazard model for default probability on Brazilian CNPJ data |
| berimbau-analytics 🔒 |
Full analytics pipeline for Berimbau Brazilian Table (NYC) — integrates Toast POS, MarginEdge food costs, ADP payroll, and Resy reservations into a unified data warehouse |
| maxsuel-rag 🔒 |
Document intelligence for a law firm — ingests PDFs, audio, video, spreadsheets; transcribes with Whisper; RAG pipeline for querying the full document base and generating legal filings |
| guios 🔒 |
Personal AI OS — multi-agent system with three-tier memory (episodic/medium-term/long-term), 22-table personal data warehouse (maps, fitness, WhatsApp, finances), Telegram interface |
| 2013–2019 Cicor Factoring |
Credit risk analyst. Six years evaluating receivables and default risk by hand. Built the statistical intuition that later became the credit scoring system. |
| 2019–2023 Duplify · co-founder |
Rebuilt the entire factoring management system from zero. XGBoost/Random Forest for credit risk on CNPJ data. Survival analysis (Kaplan-Meier, Cox model) for default time-to-event modeling. Offered a Master's position at Unicamp — declined due to bureaucracy. |
| 2023–present bee6 · CTO |
|
USP — Bachelor's in Social Sciences (FFLCH) Quantitative methods, sampling design, and fieldwork for socioeconomic research. Co-founded the Social Research Junior Institute — ran surveys and data collection for Professor Gustavo Venturi's social impact studies. The statistical foundation that later moved into finance, econometrics, and ML.