José P.
Barrantes
Currently ML Engineer at IBM, Fraud Prevention Engineering
I specialize in managing the complexity that data introduces to ML-enabled systems, and in getting those systems into production.
Day to day, I work on ML platforms, data pipelines, and the services that put a model in front of a fraud analyst. I choose the method for each question. I use gradient boosting, Bayesian analysis, causal inference, and deep learning, and some questions only need a good plot.
3 conference talks · 1 peer-reviewed paper · 900+ students trained
Experience
Eight years of building data products in fintech, cybersecurity, financial risk, and research.
-
2025 - Present
Machine Learning Engineer
IBM · Fraud Prevention Engineering · Costa RicaI lead fraud science research on patterns such as camouflage, non-payment intent, and impersonation, and we turn the findings into ML prediction services. I develop models, and most of my time goes to building and deploying ML products on IBM Cloud with fraud analysts, software engineers, and business teams.
- I designed and implemented the team's first ML platform. It runs a system that automates anti-fraud actions, which cut our reaction time and saves the company money.
- I also trained the model that serves inference for that product. More developments on the way :)
PythonMLflowDockerKubernetesFeast -
2023 - 2024
Assistant Director, Data Scientist
Moody's Analytics · Predictive Analytics ChapterThe Predictive Analytics Chapter focuses on modeling, statistical analysis, experimentation, and process automation, over large databases of climate and public financial data. We worked mostly in R, Python, and the AWS ecosystem.
- Parallelized our code and brought the execution time down, up to 30x faster.
- Experimented with several models for risk assessment and time series, and automated report generation for our customers.
RPythonAWSTime Series -
2021 - 2023
Security Data Analyst
Arkose Labs · Customer Success ChapterI protected our customers from cyberattacks. I found attack patterns in real time to identify, describe, and stop them, then met with the customers to present the findings and explain how we handled their fraudulent traffic.
- Built automated solutions for some of our most common investigation tasks.
- Our team was awarded internally several times for this work.
AWSPythonSQLStatistical Inference -
2020 - 2021
Data Scientist
Namutek Fintech · BAC Credomatic Banking GroupI ran projects end to end. I pulled data from the data lake, wrangled it, engineered features, built models, and presented the findings. Then I worked with management, product owners, and project managers to turn the findings into decisions.
- Developed and deployed more than 50 projects for the Marketing, UX, Fraud Prevention, and Compliance chapters.
- Saved budget on license costs by designing and implementing our own R Shiny dashboards.
R ShinyPythonSQLAWS -
2011 - 2018
Teaching Assistant, Biology
Universidad de Costa RicaI gave lectures, taught laboratory techniques, and evaluated students. I trained over 900 university students across biology courses.
My full history, references, and international experience are in the resume PDF.
Talks & Workshops
I present my fraud prevention and MLOps work at conferences around Latin America.
Fraud Prevention, ML & Human-in-the-Loop Design
Most fraud systems are built to take the analyst out of the loop. I propose a design where the model ranks the review queue for the analyst, and each analyst decision goes back into training as a label.
Why it matters This is how I work every day, written up as a design pattern for a regional audience of practitioners. The talk covers where full automation fails in fraud prevention, what the pattern looks like as an architecture, and how to measure if keeping humans in the loop is paying off.
Feature Store for Time-Series Forecasting
A hands-on workshop where we build a feature store for multi-series forecasting on Iowa Liquor Sales data. It covers ETL pipelines in Mage, feature definitions and point-in-time correct training datasets in Feast, DuckDB as the offline store, and online materialization to SQLite for serving.
Selected Projects
Open source work across ML engineering, data science, and analytics.
Time-Series Feature Store (Mage + Feast)
The companion codebase for my DataDay 2025 workshop in Monterrey, Mexico: a working feature store for time-series forecasting, with Mage handling the orchestration and Feast serving point-in-time correct features for training and inference.
Why it matters A feature joined at the wrong timestamp leaks future information into training, and nothing fails when it happens. This repository shows how to prevent that leakage when training and inference read the same features.
High-Performance Analytics Pipeline
An end-to-end data pipeline on Iowa Liquor Sales (30M+ rows). It ingests concurrently from the SODA API, transforms with Polars, stores the data in DuckDB, runs geospatial analytics in R, and forecasts with sktime. It comes with a hands-on tutorial article.
Why it matters The whole pipeline, from ingestion to forecasting, runs on 30M+ rows on a single machine with no cluster.
ML Model API Microservice
A microservice that predicts bike rentals with a CatBoost regression model. It has Pydantic request validation, automated unit, integration and inference tests with pytest, Docker containerization, and a CI/CD pipeline built with GitHub Actions. It comes with a tutorial article on software engineering practices for data scientists.
Why it matters I wrote this for data scientists who are moving into engineering. It applies validation, tests, containers, and CI/CD to a working model service.
Iowa Liquor Sales Analysis
An analysis of 12 years of Iowa liquor sales (~28M invoices). It uses STL decomposition, isolation forest anomaly detection with PyCaret, and Bayesian estimation (BEST) to test whether inventory diversity drives sales and whether the pandemic increased alcohol consumption. Built in R with Python interop through reticulate, and published as a reproducible Quarto document.
Why it matters Both answers come with uncertainty estimates, and anyone can re-run the Quarto document end to end.
Real-Time Fraud Prevention System
A sandbox for the triage pattern, built as five microservices in Docker Compose. A FastAPI service generates transactions with the Synthetic Data Vault (SDV), Kafka streams the events, a Mage pipeline scores the transactions, and analysts review the high-risk cases in a Streamlit viewer over DuckDB.
Why it matters I use it to test queue prioritization end to end. It is still in development.
Open Source Contribution
Open source software I maintain and use at work.
MageML
A fork of the Mage pipeline platform, extended for ML engineering. Pipelines can mix Python, SQL, R, and Rust blocks. Rust blocks run on Polars, and consecutive Rust blocks compile into one program. R blocks share one session with a locked package environment. Any pipeline can be exported as a Docker image with an HTTP API, schedules, run history, and metrics, then deployed to Kubernetes, Cloud Run, Azure Container Apps, AWS ECS, or IBM Code Engine. The same services serve inference for scikit-learn, XGBoost, LightGBM, PyTorch, and ONNX models.
Why it matters My team, Fraud Prevention Engineering at IBM, runs MageML in production on IBM Cloud. For ML work, it adds data contracts, run records, MLflow experiment tracking, model release checks, output monitoring, and a Feast client to Mage. The stack moves to Python 3.12, pandas 3, Polars 2, and NumPy 2, and MageML sends no usage statistics.
Publications & Articles
Peer-reviewed research and long-form technical writing.
-
Peer-reviewed
Specialized landing maneuvers in Spix's disk-winged bats (Thyroptera tricolor) reveal linkage between roosting ecology and landing biomechanics
High-speed videography and biomechanical analysis of landing behavior, linking roost choice to how a bat decelerates and attaches.
Read paper -
Tutorial
Building a data pipeline for high-performance analytics
How to move data from the Socrata Open Data API into a local analytical store that is fast enough to query interactively.
Read article -
Tutorial
Exposing an ML model through an API
Software engineering practices for data scientists, applied end to end to a real model service.
Read article
Skills & Technologies
The tools I reach for most. The projects above show them in use.
Domains worked in Fintech & payments · Cybersecurity · Financial risk · Retail & marketing analytics · Clinical data · Computational biology
Get in touch
I am open to collaborations and interesting problems. I also like tea, coffee, and wine.
I am based in Costa Rica (UTC−6), so my hours overlap with the US. I read every message, and I reply to the ones that are not mass mailings.