OrgForge

OrgForge simulates weeks of enterprise activity and outputs a coherent, cross-system dataset spanning the tools most corporate AI agents are expected to reason over: wikis, ticketing, chat, version control, call transcripts, customer support records, CRM data, email, and server logs. The core workflow is parameterization-first — you configure a company shape (size, industry, scenario type), run the generator locally via Docker or docker-compose, and receive a dataset where every artifact is traceable back to a defined ground truth. That ground truth is the evaluation hook: you know what the correct answer is before you ask the agent.

The differentiating feature is determinism. Most synthetic data tools produce plausible-looking noise; OrgForge produces a causally consistent simulation where a JIRA ticket references the Confluence page that references the Slack thread that triggered it. That cross-system coherence is what makes it useful for evaluating retrieval systems — you can design a question whose correct answer requires reasoning across two or three sources, then verify whether the agent found the right path.

The tool also ships with documented support for insider threat and departure cascade scenarios, described in a dedicated INSIDER_THREAT.md file in the repository. This makes it relevant for red-teaming access control logic and testing whether an agent correctly surfaces or withholds information based on simulated org-chart context. Where OrgForge breaks down is in scenarios requiring domain-specific vocabulary depth — a generated dataset for a pharmaceutical company will have the structural shape of that organization but will not replicate the precise regulatory citation patterns or product nomenclature that a real QA agent would encounter. Teams building agents for highly regulated verticals report needing to augment the synthetic output with curated real examples.

The project is MIT-licensed, ships with a Dockerfile and docker-compose for self-hosted use, and carries no API dependency, so the entire generation pipeline runs air-gapped — relevant for teams whose data governance rules prohibit sending even synthetic org data to external services.