One API call returns a complete, internally consistent synthetic persona. Sample a thousand and the population matches real demographic baselines. Pass one to a language model and it answers in character.
Married, two kids. Commutes by car to a small office in Lexington. Watches basketball most weekends.
Synthetic personas are a recent category. Here's the working definition.
A synthetic persona is a constructed individual. Each trait is sampled from real demographic data, conditional on the traits drawn before it. The dimensions span age, country, occupation, income, beliefs, lifestyle, and physical features. The output is a coherent, complete record for someone who doesn't exist.
Pass one to a language model and it answers in character. A thousand of them stand in for a real population, queryable for the cost of an API call.
Now, what to do with one. Or a thousand.
Same three steps. Generate, run through your model, read what comes back.
One API call returns a complete persona, detailed enough to use on its own. Call it again for as many more as you need. The sample stays balanced across age, region, education, occupation, and every other dimension on the record.
Pass the persona record into your LLM prompt. Ask anything: a product question, a message, a survey. The persona answers in character. Run it once for a quote, or across the sample for a pattern.
One persona gives you a story you can quote. A larger sample lets you measure patterns: splits across demographics, the overall shape of responses.
One answer is a quote. A thousand is a finding.
Generation runs in order. Fixed traits come first; later traits inherit those constraints.
Filters stack. The person still makes sense.
Most generators manage one or the other. A 19-year-old surgeon is statistically possible. Ten thousand random people skewed toward the same job aren't a population. PersonaGen has to hold both ends at once.
Traits are sampled in dependency order: the country sets up the region, the region sets up the schooling, the schooling sets up the job. Nothing in the record contradicts anything else.
One person
As the sample grows, the mix converges to the source data: age, region, occupation, every other dimension PersonaGen samples. By a few thousand draws, the sample is statistically indistinguishable from the target population.
At scale · < 2% drift
The person works. The crowd works too.
Live call against the production API. Add filters, set a seed, or let it draw from the full country model.
Six patterns we see most often.
Ask the same question across a thousand personas. Compare the answers.
Check how a model answers a new parent, a retiree, or a recent graduate.
Try a flow, message, or price with a varied sample.
Controlled cohorts for teaching, agent-based simulation, or counterfactuals.
Stories, games, or NPCs drawn from a real demographic spread.
Pass persona context to image models for more varied character sets.
Free for academic, non-profit, and open-source projects. Commercial use is paused. Send a short note about what you're building.
Large samples are intended to stay close to population-level patterns, especially for census-backed demographic fields. A single persona is a plausible synthetic composition, not a record matched to a source row.