Test-data generation (PII-safe)
The problem
Teams copy production data (risky) or hand-make 5 dummy rows (weak coverage).
Workflow
- Schemafields + rules
- LLMvalid + invalid + edge rows
- Validatorcode checks rules
- DatasetCSV / JSON / SQL
- Test rundata-driven tests
Real-world example
For a KYC form: valid PAN formats, invalid PAN (wrong length, lowercase), names with initials and apostrophes, DOB exactly 18 years ago today, pincodes for different states — all fake, none from production.
Starter prompt
Create {n} rows of synthetic test data for this schema. Include 60% valid, 25% invalid (label the
rule broken), 15% edge cases. Use realistic Indian formats. Never use real people's data. Return
CSV.
SCHEMA: {schema}Tools
LLM, Python validator, Faker
Human check
Validate rows with code — LLMs miscount lengths and checksums.
Measure success
Data-prep time; defects found via edge-case data; zero production data in test envs.