Research & Teaching

My research asks a practical question: what would it take to make trustworthy behavior part of an AI system's architecture, rather than something checked after the fact?

Publications & Presentations

"Verification-by-Design: A Security Architecture for AI-Enabled Decision Environments." Connections: The Quarterly Journal, Special Issue on AI and Emerging Security Issues (forthcoming).

"Verification-by-Design: Embedding Assurance into AI-Enabled Decision Architectures." Finalist paper and presentation, AI Assurance & Governance Summit, Stanford University, October 2026.

Memory-centric, dual-process AI architecture for longitudinal managerial judgment. Presentation, ICAD 2026 (International Conference on Axiomatic Design), MIT, June 2026.

Compliance-by-design generative AI at the edge: A memory-centric architecture for managerial decision support and organizational governance. Poster presentation, BIG.AI@MIT 2026 (Business Implications of Generative AI Conference), MIT Initiative on the Digital Economy, April 2026.

Panelist: Intellectual Property and Contracting with AI Providers. WLW (Women in a Legal World) AI and IP series, Barcelona, October 2026.

"The AI Energy Debate Is Asking the Wrong Question." Article, April 2026.

Master's thesis: AI/ML in edge computing and IoT security, University of Chicago, 2025.

Executive Education

MIT Professional Education

University of Cambridge, ASCEND Program

Cornell University, Chief AI Officer Program and AI Blueprint

Doctoral Research

Doctor of Technology (AI) candidate, Purdue University.

My dissertation develops the Structured Language Model Architecture (SLMA), a memory-centric runtime for AI used in regulated decision-making. Most enterprise AI is stateless: a model generates an answer, the answer is delivered, and the reasoning behind it is gone. When a regulator, an auditor, or a board later asks how a decision was reached, there is little to examine. SLMA is built on the opposite premise. It is a governed-cognition architecture: a runtime that bounds, logs, and commits AI reasoning to institutional memory.

In technical terms, the architecture is dual-process. A fast path handles routine inference, and a deliberative path checks reasoning before it is allowed to shape a consequential decision. Writes to persistent memory pass through an operating-system-level runtime that validates reasoning before any state is committed, so the system does not accumulate conclusions it cannot justify, and committed state carries the provenance needed to reconstruct a recommendation after the fact. Above that runtime, hypervisor-style orchestration coordinates multiple governed reasoning instances under a single governance policy, with attested communication between layers and defined points at which a decision escalates to a human.

Methodologically, the work is grounded in Axiomatic Design. Each governance property is derived from an explicit functional requirement, so assurance can be traced to a design decision rather than asserted after deployment. The architecture is evaluated against decision scenarios drawn from banking, insurance, and government advisory, and through large-scale agent-based simulation.

The work is patent-pending: nine invention families filed through the Purdue Research Foundation, which lists the SLMA Governed Cognition Architecture as available for license.