Traditional surveillance is retrospective by construction: someone must fall ill, seek care, be diagnosed, and be reported before a case exists. That chain takes weeks. Search queries, social posts, and digital behaviour move in hours — and under the right methods they carry real epidemiologic signal.
I build and validate those systems. My published work spans COVID-19 search surveillance, global infodemic analysis, vaccine safety signal tracking, and variant monitoring through query data. I also know where this field oversells itself: digital signals are noisy, confounded by media coverage, and easy to misread. The discipline is in separating the signal from the news cycle.
Most digital surveillance projects fail the same way: a correlation is found in historical data, declared predictive, then degrades quietly once deployed. The work that matters is validation against a real epidemiologic endpoint, and monitoring that detects decay rather than discovering it later.
Monitoring built on digital data streams, validated against conventional indicators so you know what the signal is actually tracking.
Measuring how health information and misinformation move through populations, and what that movement does to behaviour and uptake.
Detecting change early enough to act, while controlling the false alarms that destroy trust in a system.
Applying large language models and agentic pipelines to epidemiologic monitoring — with the validation such systems usually skip.
We define what you need to detect, how fast, and what counts as an actionable signal — before any data is pulled.
Data sources are assessed for coverage and bias, and the signal is validated against a conventional epidemiologic endpoint.
The system is built with documented methods and reproducible code, tuned to an agreed false-positive tolerance.
Monitoring for performance decay on a defined review cadence, because a surveillance system left alone stops working.
You don't need a formal brief to start. Describe what you are trying to detect and I'll tell you honestly whether digital data can detect it, and what it would take to trust the result.
If it isn't a fit, I'll say so directly — and where I can, point you to someone better placed. There's no obligation in asking.