01 · Data engineering · 2026
Radar Prospection France
Daily pipeline that detects business signals of French companies from BODACC open data. Medallion architecture on DuckDB with dbt, orchestrated by Dagster as day-partitioned assets — every day replayable on its own.
Radar reads BODACC — the official bulletin of French commercial announcements — every morning and keeps one signal: companies that just raised capital. It answers two questions: where to prospect, with the number of companies per département over the last thirty days, and who to call, with one row per company linked to its original announcement.
It is a medallion architecture on DuckDB. Bronze keeps each day's export exactly as the API returned it, silver cleans and selects, gold answers the business questions. Transformations are dbt models; ingestion and dbt run as a single Dagster asset graph, partitioned by day in Europe/Paris, so any day can be replayed on its own.
Seven dbt tests run on every build, including a uniqueness test on SIREN that proves the grain of the call list. The known limits are written down rather than hidden: the JSON bronze layer grows by about 19 MB a day, silver and gold are rebuilt in full, and the schedule only ticks while the machine running it is on.