Johnattan's Laboratory
← All experiments

An AI that read Latin American health news every morning

I built a pipeline that collected, sorted and summarised health system news from 33 countries. Later I moved it to on demand, and then I retired it.

  • News
  • Automation

Updated20 Sep 2026

The problem

I follow health systems across 33 Latin American and Caribbean countries. Reading every country's health news each morning is not possible for one person.

What was out there

General news alerts return a lot of noise, and they have no idea which stories matter for a health system.

What I built

A pipeline that collected health news from 33 countries, picked the articles that mattered, summarised them and kept an archive of events.

Top of the LAC Health Watch number 5 digest for the week to 29 August 2026, with 12 developments. It lists three signals (Colombia's insurer crisis, a rise in suicides reported in Argentina, a national doctors' strike in Paraguay) and the first two development cards, each tagged by country and trend with What changed and Why it matters lines and links to the news outlets.
One weekly digest: the pipeline grouped the week's articles into developments and wrote what changed and why it matters.
Top of an on-demand dossier titled Scarcity and shortage of medicines in Latin America and the Caribbean, dated 4 August 2026 with 57 events, showing a synthesis of five bullet points, each followed by links to the news outlets that support it.
An on-demand dossier: one question asked of the archive, answered with a synthesis that cites its sources line by line.

Equipment

Results

On 600 replayed articles, TypeSafe's Jev was better than my GPT-5.4 nano triage gate at deciding which articles mattered: an F1 score of 0.969 against 0.901.

What I learned

A daily feed is only worth its cost if someone reads it every day. Running it on demand, when I had a clear question, worked better.

A small typed judgment was better than a general model at picking the articles that matter.

Lab notes

  1. I retired the monitor and kept its archive of articles and events.

  2. I tested TypeSafe's Jev against my GPT-5.4 nano triage gate on 600 replayed articles, and Jev scored an F1 of 0.969 against 0.901.

  3. I switched it from scheduled runs to on demand, because the daily and weekly jobs cost more than the reading they produced.

  4. I split the news pipeline out of my Substack writing project and made it its own tool, covering 33 Latin American and Caribbean countries.

© 2026 Johnattan García-Ruiz Last updated