Applied AI
meddies-faker: Synthetic People from Real Statistics
A Faker extension with country-aware providers that sample names and personal attributes from sourced census and survey statistics.
Key results
- locales
- 20
- country-aware providers
- countries with sourced statistics
- 18
- census, survey and hospital data
Problem#
Default Faker providers produce names and attributes that look plausible but don't follow how they are actually distributed in a country. For a PII detector, that matters: a model trained on unrealistic names learns the wrong patterns.
Approach#
meddies-faker extends Faker with pluggable country-aware providers that sample names and personal attributes from real-world frequency distributions, drawn from sourced census, survey and hospital statistics. It is the data engine for version 3 of the Meddies-PII pipeline.
Status#
Internal today, with open-sourcing planned. The public showcase page is generated from the library's benchmark output; country data is CC BY-SA 4.0 with per-source attribution.
Limitations#
Matching population-level frequencies does not make any individual record realistic, and the statistics are only as good — and as current — as their sources.