Research
My research asks a practical question: how should organizations adopt AI systems that are powerful, cheap, and unreliable at the same time? I work on this from three directions: the economics of AI adoption, the governance of autonomous agents, and applied systems built with industry partners.
Economics of AI adoption
The Frontier TrapIn preparation
AI spending grows while the price per token falls. I argue this comes from four forces (spotlight bias, the credit illusion, Parkinson’s law of compute, and the arms race) and propose four counterweights: meter, test, cap, anchor. Based on interviews with data and AI leaders in healthcare, software, and energy. In preparation for a practitioner venue, with an academic companion paper.
Innovation costsDraft
An essay on accelerable versus scalable innovation, using the Human Genome Project and Celera as the central case.
Agentic AI: readiness, governance, security
Readiness for delegated autonomyActive
A framework for deciding when a task is ready to be handed to an AI agent, borrowing from aviation: readiness as a function of autonomy level, consequences of error, and process complexity. Includes governance and security practices (OWASP Agentic Top 10, agent governance toolkits) and the idea of human-to-agent personas for trust. Feeds MSA 8770 and the Truist bootcamp.
Applied systems with industry partners
Better Business Bureau of AtlantaActive
Over 200,000 consumer complaints. Two tracks: ChatBBB, conversational agents for complaint intake and resolution that are aware of emotion, complexity, history, and when to escalate; and a BBB analytics dashboard for cohort and archetype analysis and demand sensing from complaint streams. A journal special-issue submission is planned.
Causal inference with LLMsOngoing
IV Finder: a pipeline that reads academic papers to propose candidate instrumental variables and checks relevance and exclusion arguments. Demonstrated on the ESG-to-profitability question.
Socratic AIEarly stage
A multi-site study of guided-conversation AI tutoring.
Earlier work
- Prompt2Cypher. Natural-language to Cypher query generation for biomedical knowledge graphs (ProKinO, ICKG). With Natarajan Kannan, Lokesh.
- Supply chain knowledge graph and the CogEval benchmark. Building and evaluating a supply chain knowledge graph, with the CogEval benchmark for cognitive-style evaluation of language models.
- Kinase SNP analysis. Explainable machine learning for kinase-substrate interactions and phosphorylation prediction (Phosformer), and annotation of understudied "dark" kinases.