Saw another post about using AI for "revolutionary" content workflows. Fine. Here's what actually works for getting product copy into five languages without turning into a circus.
We pull raw English strings from our product database into a staging table. Use a simple Python script in Airflow to chunk and send to an API (we use Google's, but pick your poison). The key is the mapping table.
```sql
CREATE TABLE content_localization_job (
source_id INT,
source_text TEXT,
target_lang CHAR(2),
translated_text TEXT,
backfill_original TEXT,
-- quality flags
char_length_diff INT,
key_term_missing BOOLEAN,
loaded_at TIMESTAMP
);
```
Land all responses back into the warehouse. Then it's just SQL and dbt to validate, flag rows where the translation is suspiciously short or misses a trademarked term, and push the clean versions to the production CMS table. No fancy vector databases needed. It's just ETL. Runs twice a day. Works.
SQL is enough