Harvard and MIT team released MatrAIx, a system that creates 8.3 billion synthetic ‘users’ stitched from real-world data for product testing
AI models can now simulate human behavior in surveys, app usage, and browsing, collapsing months of user research into minutes
Same persona yields dramatically different results across models (23% to 94% purchase intent variance), raising concerns about whose voice is actually represented

MatrAIx — Simulate Before Reality

Home Research Personas Applications Playground Community

Open-source research community

Simulating the World

with 8.3 Billion Persona Agents

MatrAIx is the simulated-user evaluation infrastructure for digital products and AI systems, grounded in a population of 8.3 billion persona agents.

Explore MatrAIx →

Live population · 8,300,000,000 agents

THE MATRAIX THESIS

The next frontier for agents is understanding humanity—and learning to behave like us.

01 · WHOPersona

→

02 · WHERE / HOWEnvironment

→

03 · WHAT / WHYApplication

→

OUTPUTEvaluation Results

01 — APPLICATION TASKS

Four types of user simulation tasks.

Type 1Survey Type 2Chatbot Type 3Web Type 4App

MARKET RESEARCH · PROMOTION TEST “Which offer would make you buy today?”

Free ship

82%

10% off

74%

Returns

61%

2× points

45%

CUSTOMER SERVICE · PRODUCT RETURN

These shoes don’t fit. Can I return them?

Yes. Were they worn outside?

No, only tried on.

Your return is approved. Here’s the label. ✓

shop.test/checkout

SHOPPING WEBSITE · CHECKOUT

▲

SOCIAL APP · CREATE A GROUP

Survey

// questionnaires & feedback

Collect structured and open-ended user feedback for market research, concept testing, and preference analysis.

Chatbot

// conversational AI evaluation

Evaluate AI chatbots across task completion, customer satisfaction, helpfulness, safety, and multi-turn reliability.

Web

// web prototype evaluation

Evaluate web prototypes and features across usability, presentation, navigation, latency sensitivity, and task completion.

App

// app product evaluation

Evaluate app features and workflows across functionality, responsiveness, task success, and user preference.

02 — WHAT MATRAIX PROVIDES

From simulated users to actionable evaluation results.

Custom personas

Evaluation infrastructure

Evaluation reports

Telemetry data

Our audio model had no real sense of its users, so everyone got the same template. MatrAIx personas made every response nuanced and specific.

Unicorn Audio AI Company

I was maintaining my own eval harness instead of doing research. MatrAIx evaluation infrastructure turned a thousand sessions into a config file.

MIT PhD Candidate

One score tells me the model got worse. MatrAIx reports tell me which users it got worse for.

Harvard Postdoc

The bottleneck for our RL work is data, not compute. MatrAIx telemetry gives us real interaction trajectories worth training on.

Frontier Lab Post-training Team Lead

03 — DEMO

See MatrAIx in action.

04 — EVAL INFRA

Start exploring the simulation stack.

◉ Matr_AI_x Eval InfraPERSONAENVEVAL

8.3Bpersona corpus

1,290persona attributes

1,000+applications

Explore playground →

[

↗

Research questions, methods, and findings.

Explore research →](/research/)[

↗

Tasks, scenarios, and product evaluation.

Commerce & RetailSoftware & Developer ToolsFinancial ServicesHealthcare & Digital Health

Explore applications →](/applications/)

05 — OPEN RESEARCH COMMUNITY

Build the human layer of intelligence with us.

Join the Persona, Environment, or Application team and contribute through research, engineering, data, evaluation, or product scenarios.

Join the community →Contribute on GitHub ↗Contact us

Eight Billion Users Now Simulated