Why this is hard to rebuild

The corpus is not a dump of a single site but 73 individually hard-won integrations, each with its own pitfalls, overcome one after another:

• Portals that cut off at 2,000 results per query — solved by partitioning by day × category; discovered only after the naive corpus had silently lost 60% of the records.
• Portals that block data-centre IPs — rerouted through a residential connection.
• Session-bound pagination, ISO-8859-1 encodings, APIs whose sort order lies, feeds that republish on every status change.
• Archives back to 2002 — parts of which the sources themselves no longer serve. That history cannot be re-acquired at any price.

Rebuilding this is not merely a question of budget, but 18 months of sequential discovery — each portal reveals its traps only in live operation.

How it runs

The entire chain of collecting, normalising, translating, verifying and publishing runs autonomously — more than 50 scheduled services with failure logging, read models that rebuild themselves nightly, a signal desk written by machines and cross-checked by machines, and a public live feed covering its own operation. Infrastructure cost: under €100 a month, everything included. Marginal cost of the next country: one integration, not a team.

Structure, not mere storage

Buyers normalised across 19.3M notices; award decisions linked to registry-listed companies (NIF, SIREN, CUI, EIK, Y-tunnus …); consistent CPV classification throughout; selection criteria structured for 87% of open above-threshold procedures; title translation as a dedicated layer; activity cycles per buyer, ready for forecasting.