Paulo Azevedo

The only data hire at a ~15-person global edtech (online courses sold worldwide). The legacy setup was a PostgreSQL stuck on a small VM ("DataFy"), wired to the production databases, and it couldn't take on more data or support complex analysis. I built a platform from scratch in its place: BigQuery as the warehouse, Cloud Run services ingesting every service database, and the marketing sources (Google Analytics, Search Console, Ads) unified alongside. Scalable and manageable, with every data source in one place, built end to end and solo in under a month.

• Replaced the legacy "DataFy" (Postgres on a VM that couldn't scale) with a BigQuery warehouse, fed by Cloud Run ingestion services pulling every service database
• Unified the marketing sources in the warehouse (Google Analytics, Search Console and Google Ads), with Metabase dashboards and Slack monitoring alerts
• Mapped market penetration across all 227 countries: where Elevify already sells, where penetration is strong, and how much of the market is still open · 227 countries
• Built B2B lead scoring: flagged corporate accounts in the base (by email and professional course usage) so Elevify could offer its Exclusive B2B plan to companies certifying their staff
• Lifted the pay rate by 18% by finding the best payment option per country and switching to it, so more attempted payments actually go through · +18% pay rate
• Ran seasonality analysis tracking growth against global demand for education; the platform now sits behind ~300k course enrollments a month · ~300k enrollments/mo

Credit-risk specialist on a four-person team (executive manager, me, a senior and a mid-level analyst) at a mobile-payments fintech competing with Brazil's biggest. I owned the credit-granting and scoring models for their entire personal-loan portfolio, and rebuilt how those models were built, validated and served.

• Turned a chronically loss-making personal-loan portfolio profitable. My models replaced the incumbents across the full audience, for an audited gain of about R$1.5M per month in additional profit. · +R$1.5M/mo
• Built 10+ credit models from scratch, using internal data, Open Finance and the Central Bank's SCR registry, usually combined with a bureau score. · 10+ models
• Led the methodology to evaluate around 13 credit bureaus (Serasa, Boa Vista and 11 others) and pick the best one to integrate in-house.
• Reworked the modeling approach: auxiliary models, each capturing the best signal in one dimension, combined by a logistic-regression ensemble. More of the data modeled, fewer inconsistencies, much better performance.
• Moved the whole modeling stack to Databricks and MLflow (features, training, artifacts), and built a custom "manual inference" pipeline to serve models with around 400 features, which MLflow could not handle natively, by separating and optimizing each step. · ~400 features
• Validated the new models with a staged, ROI-backed rollout: 10% of loans, then 30%, then 100% of the audience.
• Mentored the senior and mid-level analysts and helped scale their work, and brought more lightness and calm to the team's routine.

Almost nine years at Brazil's leader in AI-based fraud prevention, from intern (while finishing my statistics degree at UNICAMP) to one of its principal data scientists. I built the core models behind the company's two main fraud lines, transactional and account-opening, and became the main technical reference for the modeling teams. ClearSale was later acquired by Serasa for R$2 billion.

• Developed the main models behind ClearSale's two core fraud lines, transactional and account-opening. These were the algorithms running the company's flagship products.
• My last anti-fraud model at ClearSale launched a major global marketplace in Brazil. They were entering the market for the first time, with no local operation and no fraud history to learn from, so I built a custom model almost entirely in the dark, on the limited data integration they could send. It launched as a success and made their Brazilian anti-fraud operation viable. · cold start
• Built a weekly engine that captured emerging fraud patterns and combined models by fraud dimension per client, raising approval rates while keeping chargebacks (CBK) under control. Brazil's largest e-commerces ran on it.
• Created a fraud-pattern detector based on a Cartesian-product grouping, with its own statistics and score per group, to surface fraud that traditional models missed. In one test, with about 100 validation SMS, I caught R$50k of fraud in a single pattern in one day. · R$50k
• Generated automatic decision rules from random forests, surfacing the most relevant ones to route about 25% more transactions through the automated path and take load off the more expensive manual operation. To run it where the data lived, I implemented the random forest itself in SQL, working across Business Analytics and Operations to land the new flow. · +25%
• Built a trustworthiness and first-party-fraud score that flagged risky orders even when we could reach and verify the real owner of the transaction. Running inside the secure authentication flow, it eliminated about 75% of the chargebacks hiding in what looked like the safe zone. It was unconventional, built on an uncommon target variable, and it paid off. · −75% CBK
• Built the company's feature platform: I migrated every model and feature into the data warehouse, so teams could reuse features, backtest and deploy models, and browse a catalog. Backtests that used to take a week ran in two hours. Every modeling team in the company used it, and it later evolved to use AI for feature generation and discovery. · 1 week → 2h
• Built the phone-ranking model for the telemarketing operation, picking the best number to reach a customer whose purchase was being verified. It cut the operation's cost by 30%+, helped control chargebacks and lift approvals, and held up so well that a later review found almost nothing to improve. · −30%
• Mentored everyone across two ~5-person modeling squads and set the technical direction; the teams' work depended on my reviews and guidance.
• Helped launch features for the new credit and insurance operations, and built the first models for the international expansion, later handed off to a dedicated team.