Senior systems and data engineer working on AI evaluation, model understanding, and reliable AI systems.

I have spent 25+ years building production software, data, analytics, and enterprise systems. My current work investigates how AI systems represent their situation, why they behave as they do, and how those behaviors can be evaluated against evidence and constraints.

Authorization Behavior Under Competing Context

Studied how language models resolve conflicts between explicit authorization policies and competing contextual information. Evaluated seven models across 100 scenarios and 11 conditions (7,700 decisions), revealing distinct susceptibility profiles and directional differences in authorization behavior.

Governed Natural Language to SQL Generation: A Behavioral and Mechanistic Analysis

Investigated how language models encode task and constraint information, combining activation and attention analysis with behavioral evaluation across governed NL-to-SQL settings.

World-state reasoning in intelligent systems

Exploring how intelligent systems represent changing state, establish which sources of information are authoritative, and distinguish errors in state understanding from errors in planning or action selection.

  1. 01

    When does successful model behavior conceal an incorrect or inconsistent understanding of its situation?

  2. 02

    How can we determine why an AI system took a particular action, rather than merely whether the action was correct?

  3. 03

    Which parts of an intelligent system’s world state should be learned, inferred from context, retrieved through tools, or represented explicitly?

My background spans software engineering, data systems, distributed systems, analytics, geospatial computing, and enterprise architecture. I am now using that experience to develop a more empirical research program around AI systems, evaluation, and model behavior.