Research that starts with how experts reason.

People do not weigh all the evidence at once. They attend to a few things, recall similar cases and fall back on familiar strategies when time is short. Cognitive architectures in the tradition of ACT-R and Soar describe attention, memory and decision strategies as explicit, testable models. We integrate that tradition with modern machine learning.

Learned representations are good at language, perception and similarity. Explicit rules and knowledge graphs are good at constraints, policy and accountability. Each fails where the other works. We study how to combine them, so that a system learns from data and is still held to the rules.

A request is only the visible part of what an expert means. Behind it sit goals, constraints and preferences that rarely get written down. We study how these become representations a system can follow, check against and revise, so that intent reaches the system as more than a prompt.

An explanation is only useful if the person reading it can use it. Many methods describe the model. Few describe the decision in the terms a reviewer already thinks in. We design explanations from how experts actually reason, and we test them with real users, not by how convincing they look.

Every expert and every team works a little differently, and retraining a system for each of them is slow and costly. Meta-learning asks another question: can a system learn how to adapt, so that a few examples of a new expert or team are enough?

Model accuracy is one number. It does not say whether the team decides well. We evaluate the expert and the system together: accuracy, calibration of reliance (does the person rely on the system when it is right, and question it when it is wrong) and time to decision.