Researcher and builder in human-centered AI.

Organizational Intelligence Loop (OIL)

• Question Enterprise AI plateaus because it lacks a grounded representation of the organization it operates in. What socio-technical layer is needed for an AI system to use organizational context responsibly?
• Method An 86-page Columbia IKNS manuscript that converts research in organizational networks, knowledge management, and AI adoption into an implementation-ready framework for enterprise AI. It treats nonresponse and behavioral silence as meaningful context about work patterns, decision flow, and trust.
• Evidence The framework defines five constructs: trust density, knowledge latency, interaction stability, nonresponse-aware signal integration, and cross-signal confidence. These measure confirmed trust, the observed speed of help and decisions, relationship continuity, patterned silence, and confidence when separate signals agree. Its reporting layer favors aggregate patterns and expertise maps, which helps teams assess contribution fairly across roles and communication norms.
• Next step OIL provides a practical, fairness-centered language for organizational context. It became the foundation for the Large Behavior Model below.

Large Behavior Model

• Question The org chart explains authority. It does not explain how work actually moves. As companies make operations queryable by AI, dumping meetings, documents, and email into retrieval misses the behavioral facts agents need: which document is treated as true, who teams actually rely on, who owns the latest answer, and where key-person risk sits. LBM asks how to read that operating layer from behavior, with privacy protection and bias controls built in.
• Method LBM is a foundation model for organizational behavior, built from scratch and currently an MVP. It learns one reusable representation of people and teams from permissioned metadata, not message content, using real and simulated company activity. Its validation holds out companies and time periods, with privacy and bias controls throughout.
• Evidence On held-out organizations, LBM's Influence Gap identifies earned influence (the expert a team actually relies on) over positional influence at 0.885, while network-structure metrics stay at chance (0.500). It predicts document source-of-truth at 0.88 AUC and attrition early warning at 0.94 AUC. On the public Enron corpus, held out with no outcome labels, it ranks the later-indicted insiders at 2.12x the base rate. In a never-seen 900-person technology company, from calendar and Slack activity alone, it rebuilt the operating map: about 70 go-to experts, about 100 hidden experts whose expertise shows up in work patterns rather than titles, about 35 key-person risks, about 70 percent backup coverage, and about 25 critical roles with no backup. Other reads, such as expertise and informal influence, are promising and still improving as training continues.
• Next step Cross-company transfer is the promising result: frozen representations pretrained across nine real companies already predict seniority in a tenth it never saw at about r = 0.50 (about 0.25 net of activity volume). The next step is to deepen it with design partners, testing on real organizations and training across more companies.

Multi-Agent Orchestration Testbed

A multi-agent system I designed, built, and run, used as a testbed for agent coordination, failure modes, and human oversight: where agents drift over long horizons, what monitoring catches it before it fails silently, and what belongs in deterministic code versus model reasoning.

Code and README on GitHub →