Org-Sim: Digital Twins from Meeting Transcripts

We built a browser-based organizational simulation system that converts recorded meeting transcripts into behavioral agent models and renders their interactions in real-time 3D. Org-Sim processes 800+ meeting transcripts, generates 49 digital twin agents representing real team members, and runs forward simulations to show how structural changes — a hiring decision, a reorganization, a key departure — will ripple through the team before those changes are made.

The entire pipeline runs locally. No backend server, no third-party data upload. All transcript processing, simulation, and rendering happens in the browser. Meeting transcripts contain sensitive organizational information; the architecture treats this as a first-order constraint.

Why digital twins over static org charts

Traditional org design relies on intuition, analogy, and retrospective reporting. Static org charts describe hierarchy. They say nothing about how communication actually flows, who decisions actually route through, or whose calendar absorbs escalations when someone is absent.

Digital twins built from transcript data describe actual behavior. HR-defined roles describe how people are supposed to behave. Meeting transcripts capture how they do behave — who defers to whom, who drives decisions, who is present in the most high-stakes conversations. Each agent profile in Org-Sim is a behavioral fingerprint extracted from observable data, not a job description.

This enables concrete scenario analysis: departure planning (remove the agent and observe which decisions now escalate), hiring decisions (simulate two junior engineers vs. one senior against the existing collaboration network), and restructuring (model communication overhead introduced by a new team boundary).

How the simulation engine works

The pipeline has five stages:

The behavioral extraction step produces per-person metrics: speaking-time ratios, question frequency, decision keyword density, interaction co-occurrence. These normalize into five traits per agent — decisionAuthority, communicationFrequency, technicalDepth, collaborationWillingness, stressTolerance — each scored 0–100.

The simulation is stochastic. Each agent generates events probabilistically from their extracted traits:

meeting_probability  = communicationFrequency × 0.2
decision_probability = decisionAuthority × 0.3

A decision with complexity > 8 and authority < 0.7 produces escalation. High-authority agents resolve decisions independently; low-authority agents facing complex problems reliably escalate — matching the behavioral data that generated their profiles.

The bottleneck scoring formula combines five signals:

score = (blocked_decisions × 15)
      + (escalations_received × 10)
      + (max(0, meeting_hours − 4) × 5)
      + (fatigue_level × 5)
      + ((10 − satisfaction_level) × 3)

In the Three.js visualization, agent nodes are positioned by a continuous Newtonian force model: preferred collaborators attract, non-collaborators repel, and a weak centering force keeps agents in frame. The spatial layout that emerges is a live social network graph — clusters signal strong collaboration subgraphs, isolated nodes signal integration risk.

Agent traits are editable at runtime. Changing an agent's decisionAuthority slider during a running simulation immediately re-initializes the engine with the new profile — turning the tool from a run-and-observe system into an interactive organizational laboratory.

Where it applies

Org-Sim is Worksona's answer to "how do we measure the impact of organizational changes?" It sits in the Domain Specialization layer of the portfolio — applying multi-agent modeling infrastructure to organizational behavior analysis. Leaders who need evidence before a restructuring, people managers evaluating attrition risk, and researchers studying team dynamics are all direct users.