When the narrative on security
becomes queryable data.

In Santa Marta, as in most Colombian cities, information on security is published every day — but in narrative language, fragmented across dozens of outlets, and never consolidated.

For journalists, authorities, and researchers, this means that structural questions cannot be answered with existing coverage. How many incidents this month? In which neighborhoods? What types? How many notes cover the same incident? Each answer requires weeks of manual reading.

Official statistics exist, but they arrive late and aggregated. What was missing was not a reporting system nor a replica of statistics — it was a quantitative reading of the media narrative itself, with auditable data quality.

We designed a deterministic pipeline that goes from raw HTML to canonical geolocated event, with an audit trace at every step:

1. Ingestion — scraping and RSS over the main regional sources, with duplicate detection and editorial windows.
2. Classification — each note passes through an LLM (Anthropic, Gemini, OpenRouter, or Ollama) that extracts 32 categories, severity, event date, and 6 enrichment enums in a single call.
3. Deterministic geolocation — v3 model that resolves event location with a clue-priority table, validates coordinate–neighborhood coherence, and leaves geo_audit per event.
4. Consolidation — a resolver with 5 criteria fuses multiple notes into a canonical event, preserving traceability to each source.
5. Extractors — over each event, victims, artifacts, judicial milestones, and recurring actors are detected.
6. Visualization — Leaflet map as permanent substrate + government dashboard + /investigate agent.

The entire stack is inheritable: if KOHR disappeared tomorrow, a new firm could maintain Sereno. That is the main filter.

See the product page: Sereno → · Live demo: serenocol.vercel.app →