inferred from GitHub
audiencetwin
AudienceTwin — test your campaign on the whole city before spending a dollar. Synthetic market from real US Census PUMS microdata.
Core trick
Samples a synthetic population from Census PUMS microdata using survey weights, then has it react to ad copy to produce weighted intent funnels.
Evidence
README matched "Claude Build Day": [Tanmayi Dasari](https://tanmayi2.github.io/) (2nd place, Claude Build Day, June 2026),
Tags
README excerpt
# audience twin
**test your campaign on the whole city — before spending a dollar**
AudienceTwin is a synthetic market for testing marketing before it ships. Paste ad copy, a
tagline, or a launch announcement, and a population sampled from **real US Census PUMS
microdata** — weighted like a proper survey — reacts to it:
- **Funnel** — attention → relevance → intent, weighted, with a 95% CI on intent.
- **Audience builder** — compose targets the way ad platforms do (age, income, education,
housing, occupation, survey-derived interests) with a live census-weighted reach count;
every test then runs on exactly that group.
- **Segments** — who it excites and who it repels, with the *why* in each archetype's own words.
- **A/B/n arena** — paste 2–5 variants (one per line) for an instant ranked comparison,
including where the ranking flips by demographic.
- **Targeting export** — one click turns the responding segments into an ad-platform-ready
targeting spec (age/income/education/tenure/geo filters + a media brief).
- **Counterfactual shocks** — branch the simulation, broadcast an event ("competitor cuts
prices", "brand hit by bad press"), and re-test to see the resonance delta.
**Privacy story: nobody is tracked.** There is no user data anywhere in the system — the
market is rebuilt from public, anonymized census microdata and public survey statistics
(ATUS routines, CEX spending, Pew religion), then brought to life with seeded, deterministic
personas. Segment-level