Sell a weekly digest of Spanish public tenders to small fire-protection installers and maintainers. The hypothesis: generic tools miss tenders that are hidden inside broad maintenance contracts, and a niche service with fewer false positives would be worth a monthly subscription.
- 436 contract folders carried a fire-protection CPV (215 still open), with a median open budget of 94k EUR (state platform) and 184k EUR (aggregated regional platforms).
- About 215 more open tenders mentioned fire-protection work without a fire CPV, usually inside "mantenimiento integral" contracts. Roughly half of them were noise: fire engines, firefighter gear, forest prevention.
- Real feeds keep old folders in "published" status for years, so a useful digest must drop past deadlines.
Details: data-sources.md.
A working pipeline in about a day, with tests on real fixtures: ingestion, CODICE parsing, explainable scoring, SQLite state and digests. Run on real data, it ingested 73,030 tenders and produced regional digests with 16 to 41 relevant tenders each (samples).
A market check, done after building, showed the need was already covered:
| Option | Price (October 2026) | Covers |
|---|---|---|
| PLACSP subscriptions (official) | Free since February 2026 | Email alerts across the whole Spanish public sector, regional platforms included |
| Commercial alert services | About 20-50 EUR/month entry plans | Alerts by CPV, keyword, region and budget; several add AI summaries of the specifications and award/competitor intelligence |
A niche digest with better precision is a thin advantage against a free official alert plus inexpensive tools that also search inside the specifications. The expected price (49-149 EUR/month) was above the market for less functionality.
- Check the market before writing code. Ten minutes of searching for existing services and their prices would have reframed the project.
- Measure the data first. The volume and noise measurements were the most valuable part of the work and needed only a few scripts.
- The hard part was relevance, not ingestion. Precision on real data (exact CPV membership, context words, penalties, stale folders) took most of the iterations.
- Keep the experiment cheap and the exit clean. Standard library only, tests on real fixtures, and a written decision make the work reusable even when the business idea is dropped.