ATLAS: Simulating Public Opinion Before You Commit to a Decision

Market research panels take weeks. Social listening tools look backward. ATLAS runs forward: upload your documents, describe your scenario, and get a simulation of how hundreds of distinct agents — each with a persona, a stance, and realistic social influence — would respond across Twitter and Reddit environments.

This is ATLAS's core proposition: make public opinion dynamics computable before a decision is irreversible.

ATLAS has its own dedicated blog at /atlas with deeper technical content and development notes. This post is a portfolio-level introduction.

What It Is

ATLAS is a Python/Vue 3 application that converts real documents into a structured simulation. The workflow runs in five stages: document ingestion, knowledge graph construction, agent persona generation, dual-platform simulation execution, and ReACT-based report synthesis.

Users upload PDFs or plain text files, provide a scenario description, and ATLAS extracts entities and relationships into a Zep Cloud knowledge graph. That graph seeds the persona generator, which produces MBTI profiles, background stories, stances, communication styles, and follower counts for each simulated agent. Follower counts are drawn from entity-type-aware Pareto distributions — media outlets receive hundreds of thousands of followers; individuals receive proportionally fewer — reflecting observed social media distributions rather than uniform random assignment.

The simulation runs both Twitter and Reddit environments simultaneously using the CAMEL-AI OASIS engine. Two narrative events are injected at approximately the 24-hour and 48-hour simulation marks to test whether agent opinions are stable or susceptible to disruption. Agent actions write back to the knowledge graph in real time, so the graph becomes a living record of the simulation rather than just its static input.

Why It Matters

The decision to announce a product, publish a policy, or launch a campaign is often made with incomplete information about how it will land. Traditional alternatives — focus groups, panels, expert advisory sessions — are slow and expensive. Social listening tools tell you what happened after you acted, not what might happen before.

ATLAS addresses this gap by making the simulation both fast and transparent. The ReACT-based Report Agent does not summarize the action log in a single pass. It runs a Thought → Tool → Observation → Final Answer loop using five purpose-built tools: insight_forge for deep pattern mining, panorama_search for broad knowledge graph traversal, quick_search for targeted entity lookup, interview_agents for simulated direct interviews with individual agents, and simulation_analytics for statistical analysis of engagement and sentiment trajectories.

Every finding in the final report is traceable to a tool call and an evidence source. The reasoning is not a black box.

How It Works

The LLM interface routes all inference through a single OpenAI-compatible client. Swapping from GPT-4o to a local Ollama model requires changing two environment variables. ATLAS can run fully air-gapped with local inference and a Neo4j knowledge graph in place of Zep Cloud — the Atlas2 variant covers this deployment model for privacy-constrained environments.

Pre-built use case templates are included for new product launches, DTC funnel analysis, and M&A financial stakeholder modeling. Custom scenarios require only documents and a description.

Where It Fits in Worksona

ATLAS occupies a distinct position in the Worksona portfolio: it is the only project focused on operationalized scenario planning. While other projects help teams understand and use AI effectively in their daily workflows, ATLAS answers a specific and prior question — what happens if we do this?

The provider-agnostic LLM abstraction and knowledge-graph-as-agent-memory patterns ATLAS establishes are directly transferable to other portfolio projects. The dual-variant model (cloud-dependent ATLAS and fully-local Atlas2) demonstrates how the same workflow architecture can serve organizations with different infrastructure constraints.

For teams that want to test their communications, product launches, or policy announcements against a simulated public before committing, ATLAS provides a structured, evidence-backed answer in hours rather than weeks.