JPMorgan Chase

01 · Context

Machine learning under real accountability.

At JPMorgan Chase, I worked on machine learning systems used in credit risk, fraud detection, and compliance analytics. The work required more than predictive performance. Models had to be explainable, governed, monitored, and dependable enough to support decisions in a regulated environment.

That setting taught me to treat model risk, data quality, and operational controls as part of the core technical problem.

02 · Credit risk

Turning behavioral signals into risk insight.

I developed and deployed credit risk models using XGBoost, Python, and AWS SageMaker to score more than 1.5 million loan applications. The models improved the Gini coefficient from 0.52 to 0.61 while maintaining target non-performing-loan thresholds.

I engineered behavioral features around payroll stability, spending patterns, and savings indicators using PySpark, WOE encoding, and IV-based feature selection. SHAP-based explanations were then used to support adverse-action reporting and regulatory review.

03 · AML and fraud

Reducing noise without losing the signal.

I built AML anomaly-detection systems using Isolation Forest, autoencoders, TensorFlow, and LightGBM. The models achieved an AUROC of 0.84 and reduced false positives by 40%, helping investigators spend more time on meaningful cases.

I also created graph-based features in Apache Spark GraphX to uncover transaction relationships, fraud networks, and second-degree connections that were difficult to capture with row-level models alone.

04 · Deployment

From experimentation to reliable service.

I helped deploy scalable machine learning pipelines on AWS using S3, EC2, SageMaker, FastAPI, Docker, Kubernetes, and Airflow. These systems supported real-time inference below 300 milliseconds and monthly batch scoring across more than two million customers.

MLflow-based lifecycle management and automated drift monitoring helped make model performance auditable and stable after deployment.

05 · What stays with me

Good models must earn trust.

My time in financial services made explainability and governance practical concerns rather than abstract ideas. A model can be statistically strong and still fail if people cannot understand, monitor, or responsibly use it.

That experience continues to influence the way I build machine learning systems today.

06 · The next question

How can risk models remain rigorous without becoming opaque?

I want to keep exploring systems that balance predictive power with transparency, monitoring, and human review. In high-stakes environments, the next improvement is not always a more complex model—it may be a clearer explanation, a better control, or a decision process that knows when to pause.