Advanced analytics & AI
Verifying epidemic intelligence with an LLM
Experts were spending three days triaging noise. We put a language model in front of them so they only saw signals worth a human judgement.
Expert review: 72 hours to under 1 hour.
The challenge
Case-based surveillance is the gold standard for disease monitoring, but it only starts counting once patients reach a health centre. By then an emerging outbreak has had days or weeks to spread, and the window for early quarantine and response has closed.
The alternative is to watch the informal signals (outbreak reports, rumours, local news, social media) and have qualified health practitioners verify them. That works, but it does not scale: every incoming item needed manual reading and classification, and expert review took around 72 hours.
What we did
We built a virtual epidemic intelligence platform that aggregates signals from many sources and routes them to a worldwide community of health practitioners for verification, with a language model doing the first pass.
- Aggregation. Outbreak articles, rumours and social posts are pulled together into one queue instead of a dozen inboxes.
- LLM classification and summarisation. Incoming items are categorised and condensed automatically, so a reviewer opens a briefing rather than a raw feed.
- Humans kept in the loop. The model prioritises and prepares; practitioners still make the verification call. Accountability stays with people.
Results
Faster, scalable outbreak verification improved global situational awareness and preparedness. The same pattern applies to any expert review queue where volume, not expertise, is the bottleneck.
Have a review queue that will not scale?
Compliance checks, claims triage, document review, tender evaluation: the shape of the problem is the same. We can usually prove or kill the idea in one MVP cycle.