Skills & Capacities

Networks and relational systems

Social-network analysis has been central to my research since graduate school. I work with social, organizational, communication, collaboration, and machine-agent networks, as well as relational and dyadic data.

Current work extends these methods to networks of AI agents and to settings in which the relevant network is only partially observable.

AI systems and evaluation

My current computational work includes large-language-model agents, multi-agent systems, machine-to-machine communication experiments, AI-mediated communication, model-supported measurement, and system-level evaluation.

I build environments in which communication topology, information access, message content, routing, synthesis, model placement, authority, and other features of an AI system can be manipulated or recorded explicitly.

A recurring analytic distinction is between component capability and system performance. A system can fail because an individual model lacks the necessary capability, but it can also fail because information is poorly routed, dependent evidence is treated as independent, synthesis introduces error, or interaction among otherwise capable components changes the collective result.

AI-assisted research and development

AI-assisted research and development is also a methodological problem in its own right. I distinguish model output from tool execution, deterministic computation from model judgment, and ephemeral interaction from durable computational artifacts. I work on how these components should be combined, validated, traced, and tested when language models participate in research or software workflows.

Three arrangements are worth keeping apart. A model response can be the final artifact. A model can be a runtime component inside a larger system, where its behavior is part of what executes. Or a model can act as a development and orchestration layer that produces durable external artifacts which then run on their own. My work spans the second and third, and they carry different validation obligations.

Model capabilities, interfaces, and failure modes change rapidly enough that these workflows require continued empirical evaluation. New capabilities can change which tasks are sensible to delegate, but they do not remove the need to validate outputs, preserve provenance, test software, and distinguish model judgment from direct computation.

Simulation and computational modeling

I build agent-based and computational models when a proposed mechanism cannot be isolated cleanly in observational data or when the relevant counterfactual does not occur naturally.

I use simulation primarily to make assumptions executable. The relevant standard is not whether a simulated population resembles an observed population in every respect. It is whether the representation is sufficient to separate the mechanisms or competing explanations the model is intended to test.

Experimental design, causal inference, and adaptive experimentation

I design randomized and quasi-experimental studies across laboratory, field, survey, product, organizational, and network settings.

The relevant design depends on the inferential target. Description, prediction, treatment effects, mechanisms, and optimal intervention policies require different comparisons and impose different assumptions.

Measurement and psychometrics

A substantial part of my work concerns measurement itself.

Measurement validity is part of the research problem. Precision does not repair a measure whose observed variation does not support the interpretation assigned to it.

Statistical, longitudinal, and predictive analysis

Mechanism tests and mediation are useful where the design identifies the relevant process. Where it does not, I distinguish an explanatory interpretation from a directly tested mechanism.

Text, trace, and qualitative analysis

Some of my current work uses observable messages or traffic to infer properties of an otherwise hidden communication process. In those settings, the central question is not merely predictive performance but which features of the trace remain diagnostic as observation becomes incomplete or contaminated.

Language-model-derived measures are treated as measurements requiring validation rather than as ground truth. Depending on the task, validation can involve human judgments, conventional statistical measures, embedding-based comparisons, cross-model checks, known generating structure, behavioral outcomes, or other external criteria.

Software and research infrastructure

I build software required to conduct, measure, validate, and reproduce research.

I have also developed and distributed purchaser-facing software that underwent purchaser code-quality and security review.

AI-assisted development

My development workflow is heavily LLM-assisted and tool-mediated. I use models to work through repositories, write and refactor code, construct tests, debug systems, operate development tools, and build research software. The intended output is usually a versioned, inspectable, executable artifact rather than a model response that must be trusted as produced.

I remain responsible for architecture, statistical and computational logic, code inspection, debugging, testing, reproducibility, validation, and the behavior of the resulting system.

Development-time and runtime model use are different things. Some systems include model-based components during execution, and their behavior depends on those components. Other software is written with substantial model assistance but produces conventional executable code, analyses, simulations, pipelines, or interfaces that do not depend on reproducing the original interaction. A simulation is defined by the simulation code that runs, and a deterministic computation remains deterministic whatever helped write it.

The preferred outcome is therefore a durable computational artifact: versioned code, an executable analysis, a simulation, a validated transformation, a software application, a test suite, an analytical pipeline, or a reproducible workflow. It should be inspectable and runnable independently of the generative interaction that helped produce it.

Behavioral and organizational science

Applied research and evidence translation

Applied work often requires integrating evidence produced under different designs and for different purposes. I synthesize peer-reviewed research, experiments, behavioral and product data, surveys, relational data, qualitative evidence, operational information, and external sources to determine which conclusions are supported, which depend on additional assumptions, and where another study or measurement effort is required.

Those outputs have been produced for scientific, technical, product, organizational, policy, client, and public audiences.