Skip to main content
← All work
Case study 02

OpenModel

100,000 simulated consumers, run before a single dollar of marketing spend.

Year
2024
Role
Author & Researcher
Stack
LangChain · Python · LLMs · Statistical Evaluation · Claude Code
OpenModel

The problem.

Surveys are expensive, biased, and slow. A/B tests need real users—and real budgets—to learn anything. Product and policy teams need a faster, cheaper sandbox to pressure-test ideas before committing capital.

The approach.

Generate 100,000 LLM-backed personas grounded in real demographic and psychographic distributions. Run policy, pricing, or messaging interventions against the population. Then—the part most synthetic-data work skips—statistically evaluate alignment, consistency, and emergent behaviors so the results are interpretable instead of vibes-based.

The outcome.

A reusable simulation framework that lets product and policy teams ask 'what would 100k people probably do?' and get a reasoned, citable answer in hours—not weeks. The work formed the basis of a published thesis and an early commercial deployment.

Next project Motorola · Edge AI