Approach
We started from the real data, not from a farm: Copernicus satellite, ocean and weather products and Marine Institute buoys over Bantry Bay, Co. Cork. On that real ocean we placed six example farm profiles, mussel longlines at Gearhies and Adrigole, oysters at Whiddy, a kelp farm at Bere Island, Castletownbere harbour and Glengarriff boat tours, each with its own species thresholds and activity limits, so the product was designed around named activities from day one rather than a generic risk score. A grower's own records arrive only when they start with their own business.
The data layer came first. A dataset registry holds every product key, dataset id and DOI, and a nightly ingest chain pulls the Copernicus physics, biogeochemistry, wave and ocean-colour products, ERA5 and ECMWF weather, Sentinel-2 turbidity and kelp indices, and the Marine Institute buoy, tide-gauge and HAB bulletin feeds. A 1993 to 2022 climatology is built from the multi-year reanalysis so anomalies and marine heatwaves have a baseline.
Honesty was engineered in rather than written in a disclaimer. Conditions are favourable, marginal or unfavourable for a named activity and never 'safe'. Bloom risk is always flagged as regional and not verified for the bay. A model-skill job scores the regional model against the buoy and the operator's own logger every night, and that score caps the confidence shown on all five signals.
The assistant is grounded by construction: it answers only from the signals, evidence and graph the product holds, scores every answer for grounding before it is shown, and redacts personal data before any text reaches an AI provider. The product shipped with an AI literacy page and an AI notice to meet Articles 4 and 50 of the EU AI Act, and a model card that states the limits.
Adding a bay is configuration only: a bay is a YAML entry plus the public aquaculture site register, with no code change. Every Irish bay is ingested today, and European bays are coming soon.