Freddy Lecue

Freddy Lecue leads AI research and model methodology at Wells Fargo, where his team builds foundation models for financial services, including fraud detection. His work connects production machine learning with two longstanding research questions: how models can remain reliable when their operating conditions change, and how people can understand the decisions those models make.

As managing director and head of Frontier AI Model Methodology, Lecue works across generative AI, agentic systems, and advanced machine learning. His responsibilities include modeling standards and the model lifecycle—from development and documentation through testing and validation—with performance, latency, cost, and risk shaping production decisions. He has described his team’s foundation models as proven at the scale of billions of debit-card transactions in fraud-detection applications.

From semantic reasoning to trustworthy machine learning

Before Wells Fargo, Lecue held senior AI leadership roles at IBM Ireland, Accenture Ireland, Thales Canada, and JPMorgan Chase. His research spans knowledge representation, reasoning, explainable machine learning, and language-model alignment. At JPMorgan, he coauthored work on feature attribution, harmful distribution shifts, sequence-classification explanations, and reward-model interpretation.

These collaborations address different parts of model reliability: identifying which inputs influence a prediction, detecting changes that could undermine performance, and testing whether explanations help people reason about model behavior. Three projects illustrate how that work has developed:

• Learning when conditions change. In 2017 research on semantic concept drift, Lecue and his coauthors combined logical reasoning with supervised learning to address abrupt changes in streaming data. Their representations included both observed facts and conclusions inferred from background knowledge. This let the learning system account for whether historical conditions were consistent with, or contradicted, other snapshots rather than relying only on how recently data arrived. The team evaluated the approach on bus-delay prediction in Dublin and air-quality forecasting in Beijing.

• Testing whether interpretable models are understandable. Lecue and his collaborators challenged the assumption that publishing logistic-regression coefficients gives users an explanation they can readily use. Those coefficients describe changes in log-odds; understanding their effect on a predicted probability requires additional, nonlinear calculations. In their 2024 study of logistic models, skilled participants struggled to predict the magnitude of output changes. The team introduced Linearised Additive Models, which approximate trained logistic models without retraining and express contributions more directly in probability terms. Participants answered model-reasoning questions more accurately with these approximations. Tests on public financial datasets found small classification-performance and calibration penalties, with a further limitation: extreme probabilities can become predictions of complete certainty.

• Explaining a reward model’s preferences. His coauthored 2025 research on contrastive explanations examines why a language reward model prefers one response over another. The method generates nearby response comparisons that deliberately change manually specified evaluation attributes. These controlled variations reveal local preferences, while representative examples help researchers investigate attribute sensitivity and compare different reward models.

Across this work, explanation is something to test: can a person use it to anticipate a model’s behavior, and can researchers use it to identify consequential differences between models?