Joe Fernandez ThinkMorphic Press

I write about putting AI to work in banking and everyday jobs, and I build with it myself.

My career started in 1994, writing COBOL and PL/SQL against IBM DB2 on mainframes. Three decades later I build and deploy machine learning models, and I write books about how organizations can use AI without losing control of it.

Along the way I worked in London, India and the United States, with IBM, Infosys and Visa, and more than twenty of those years have been in financial services. I taught myself C, C++ and Python, and I went back to school for AI and machine learning at MIT, UC Berkeley Haas and Johns Hopkins.

Before AI, I had a research paper, articles and blog posts, but no books. AI changed that for me. I use it to research, to argue with and to learn from, and I want other people to see what it can do for them too. ThinkMorphic Press is where my books are published.

Books

Published by ThinkMorphic Press.

The Seatbelt Principle

AI Governance for Banking Executives Who Intend to Move Fast

Governance works like a seatbelt, the thing that lets a bank move fast with AI and still stay in control. The book is written for boards, CXOs, and legal and risk leaders, with little to no math, and each chapter builds on a discipline banks already practice, such as dual control on a vault or underwriting a loan. It closes with ninety-five questions a director can put to management, three of which can only be answered by a demonstration.

Get the Kindle edition on Amazon Author page on Amazon

A book on fraud and AI

Fraud is the first control function where AI runs on both sides at once, for the attackers and for the institutions defending against them. This book is written for executives across the C-suite, because fraud is a financial, operational, security and reputation problem all at the same time. In progress.

Projects

Machine learning work with the code on GitHub. Several of these began as coursework in the UC Berkeley Haas AI and machine learning program.

  • Pricing a used car

    Used car appraisals are subjective, which leads to weak negotiations and inventory that sits too long. I built a model to predict a fair price from eleven attributes, including brand, engine power, mileage, owner type and location, and compared nine regression models. Gradient boosted trees did best, and I recommended making the model's price the benchmark for every trade-in offer.

    XGBoost, R² on held-out data 0.958 View on GitHub

  • Movie recommendations at scale

    I compared seven collaborative filtering algorithms on a slice of the MovieLens 25M dataset holding 1.1 million ratings, using five-fold cross-validation. Matrix factorization (NMF) gave the lowest error. The neighbor-based models were accurate enough but took 20 to 80 seconds to score, so they only make sense when recommendations are computed in batches ahead of time.

    NMF, RMSE 0.849 View on GitHub

  • Diabetes risk from health indicators

    Using 253,680 records from the CDC's health indicator survey, I looked at how lifestyle factors such as diet, physical activity, smoking and BMI relate to diabetes, and analyzed men and women separately. Tuned logistic regression gave the best balance of accuracy and run time. Recall stayed low, and improving it is the next step for this work.

    Logistic regression, accuracy 0.87 View on GitHub

  • Which customers say yes to a term deposit

    A bank's telephone marketing campaigns produced 41,188 records, and the question was which customers are worth calling. I compared four classifiers. Logistic regression matched the support vector machine on accuracy while running in 0.15 seconds instead of 30. I kept call length out of the model, because a bank cannot know how long a call will last before it makes it.

    Logistic regression, accuracy 0.91, ROC-AUC 0.93 View on GitHub

More on my GitHub, including Titanic survival prediction with PySpark deep learning and weather forecasting with a recurrent neural network in Keras.

Research

Peer-reviewed work from my years at Visa.

We clustered five million alerts and 67,000 incident tickets from a large enterprise IT environment to find the groups of events that belong together, so operations teams can manage incidents instead of individual alarms.

Lin, D., Raghu, R., Ramamurthy, V., Yu, J., Radhakrishnan, R., and Fernandez, J. “Unveiling Clusters of Events for Alert and Incident Management in Large-Scale Enterprise IT.” Proceedings of KDD ’14, ACM, pp. 1630-1639. Read the paper

About

I am a senior technology leader at a large U.S. commercial bank, responsible for technology program delivery across risk, financial crime and corporate functions. My background spans anti-money laundering, fraud, sanctions, know your customer, data privacy and governance, risk and compliance, along with portfolio governance, product management and operations.

  • B.S. in Computer Science, Bangalore University
  • MBA in Technology Management, University of Phoenix
  • AI and machine learning certifications from MIT Professional Education, UC Berkeley Haas and Johns Hopkins

I live in the San Francisco Bay Area. You can write to me at [email protected] or reach me on LinkedIn.