Senior Machine Learning Engineer

Every credit decision at Facio runs on a model: who gets access, how much limit, when to reach out in collections, which offer to show. Too many of those models lose performance on the way to production, because the data they were trained on is not the data they see live. You will own the platform that closes that gap: features that are the same in training and in production, a standard way to ship and run models, and the monitoring that tells us when something drifts. 1️⃣ Features first One feature platform, consistent in training and production. 2️⃣ Then retraining Automated retraining once features are solid. 3️⃣ Then many models Several models side by side, with automatic traffic allocation. What you will do
• Build the feature platform that keeps features consistent between training and production — for real-time scoring and for heavy batch jobs over millions of users (Open Finance histories, collections attempts, transactions)
• Make it easy for data scientists to create and test new features, with no CSV files moving between notebooks and SageMaker
• Standardize how models are trained, versioned and deployed, so product engineers send a payload and get a score back
• Own model serving in production: endpoints, deploys, rollbacks, alerts, feature and model drift
• Build the infrastructure to run multiple models in production and shift traffic automatically, from A/B tests to contextual bandits
• Put Claude and agents to work inside the platform wherever they help: labeling, evaluation, tooling You are the right person if you
• Have run ML systems in production and owned what happens after deploy, not only the training
• Have solved training/serving skew before: point-in-time correct features, online and offline stores that agree
• Have built data pipelines for both low-latency serving and compute-heavy batch
• Write strong Python and are hands-on with AWS and infrastructure as code
• Know tree-based models (XGBoost, LightGBM) well enough to work side by side with the data scientists who build them. You don't need to know credit policy
• Prefer building systems that serve many models over building one model
• Have standardized the model lifecycle somewhere: registry, versioning, CI/CD and retraining, with every prediction traceable to the model, features and data behind it ➕ A plus, not a requirement: you've worked with contextual bandits or online experimentation in a live product. Our tech stackWe are an AI-first company and every team ships with AI leverage. On the ML side, this is what we use today. The tooling calls going forward, including build vs. buy, are yours to make. MODELING: Python, XGBoost and LightGBM for the credit models
· MLflow for tracking and registry TRANING: SageMaker for training and model endpoints
· Athena and S3 for offline data
· DynamoDB and Lambda for online features BAKEND: Serverless on AWS Lambda, Step Functions and DynamoDB
· infrastructure as code with AWS CDK
· fully automated CI/CD Claude and coding agents in every engineer's daily work, with internal skills and cookbooks 🌎 Fully remote We meet 4 times a year in São Paulo — flights and accommodation on us. 💰 Above-market pay CLT contract, with compensation above market. Deel for international EMPLOYEE 📈 Top-tier equity A strong stock options package. 🤝 Senior team You'll work with a talented senior engineering team.