reef.home + ReefWatch
Peter's operating reef provided the verified source for ATO and reference-temperature history connected in July 2026.
What exists today
One narrow founder-tank telemetry pilot, separated from the product simulation.
reef.home is the local address of Peter's ReefWatch installation. ReeferVision is the cloud product layer receiving an allowlisted, read-only copy of selected telemetry. The public dashboard and workflows are a separate simulation using fictional sample values.
Signed telemetry ingestion
An outbound hub keeps an unsent queue, signs each batch, and sends accepted readings through Cloud Run into BigQuery.
Vision, alerts, and workflows
The public product experience uses fictional values. Camera ingestion, models, notifications, and automation remain validation work.
1living founder reef used for validation
2cloud-connected streams in July 2026
24integration checks passing on 20 August 2026
0cloud-to-equipment control routes
Evidence boundary: this proves a working engineering path on one founder-operated system. It is not yet evidence of model accuracy, improved coral growth, continuous fault-free operation, or a biological early warning.
Technical challenge for the program
Useful multimodal prediction with a small, diverse fleet.
How do we combine sparse chemistry telemetry, equipment events, and per-colony visual histories into useful early warnings — while controlling false alarms, personalising to each tank, and keeping inference affordable?
This is where focused support from Google engineers would be most valuable: multimodal feature design, evaluation methodology, active learning, model-serving economics, and the transition from simple statistical baselines to production vision models.
Validation plan
The next twelve months are about evidence.
01
Extend the founder pilot and instrument external tanks
Add repeatable image capture, stronger device identity, and paired histories across different system types.
02
Ship useful statistical baselines first
Measure drift detection, probe cross-validation, equipment signatures, and operator response before introducing complex models.
03
Build and evaluate coral vision
Create a labelled longitudinal dataset and evaluate change detection against each colony's own history, not a generic visual threshold.
04
Test the commercial fleet workflow
Validate whether aquarium service businesses and aquaculture operators gain enough operational value to support a scalable subscription.
Why Google Cloud
Managed infrastructure lets a small team spend its time on the biological signal.
BigQuery now; Pub/Sub next
The pilot stores accepted telemetry in BigQuery. Pub/Sub is planned when independent consumers are introduced.
Vertex AI
Train and evaluate vision models only after repeatable image capture and a labelled longitudinal dataset exist.
Cloud Run now; Firebase next
Cloud Run handles pilot ingestion. Firebase is planned for authenticated live product state and notifications.