How Verily Health monitors wastewater for 30-plus pathogens with a 5-person team and 1 dbt package

At a dbt Summit 2026 breakout session titled "dbt for public health: sightline wastewater monitoring at scale," David Drummond, data engineer and technical lead at Verily Health, described how a five-person data team uses dbt to power near real-time wastewater surveillance for more than 30 pathogens.

Verily Health's public Sightline dashboard lets anyone track regional and seasonal trends for pathogens like COVID-19, influenza A, norovirus, and rotavirus, built from wastewater samples that government partners collect and Verily's labs process.

That data also reaches academic researchers, government agencies, and organizations like the CDC, typically arriving with a lag of only a day or two.

Drummond's team builds on FHIR (Fast Health Interoperability Resources), an open healthcare data standard maintained by the international HL7 organization. FHIR follows an "80/20 rule": common use cases get built into the standard itself, and everything else can be handled through extensible implementation guides.

FHIR's more than 150 resource types, things like Patient, Observation, Specimen, and Location, arrive as deeply nested JSON or XML. That creates a recurring design problem: how far to flatten the data, what to use as a primary key, and whether to track full history or just the latest snapshot.

Rather than solve that problem separately for each of the two or three dozen datasets the team supports, spanning wastewater monitoring, diabetes management, weight-loss programs, and clinical trials, Drummond's team built a shared dbt package with parameterized macros for each FHIR resource.

The same observation macro that tracks influenza A levels in a wastewater sample also tracks a person's A1C reading from a continuous glucose monitor. Sensitivity and grain get adjusted through macro parameters rather than duplicated code.

Two heuristics keep the resulting models usable: composite keys are capped at around three columns, and most tables stay at roughly 30 columns or fewer, even though that means choosing a grain other than the finest one technically possible.

The pipeline follows a medallion structure: a restricted, permissioned bronze layer feeds source-aligned staging models that mirror FHIR terminology, optional intermediate models for testing, and domain-centric mart models built in language researchers and scientists actually use.

dbt's freshness tests monitor not just the team's own pipeline but the wide range of formats, from fixed-width files to CSVs, that arrive from academic partners.

Because the team is small relative to its stakeholder base, Drummond said the group leans on dbt's semantic layer and documentation, including dbt docs, to support self-serve analysis rather than acting as a bottleneck.

The approach was tested at scale this past summer, when Verily partnered with health agencies and academic institutions to extend Sightline monitoring to World Cup host cities, including the stadium in Kansas City. Drummond reported no major outbreaks during the tournament.

Watch every keynote and breakout from dbt Summit 2026 on demand. Watch dbt Summit 2026 on demand →