Case study · 01 of 04
AGRIGRAPH AI
Forecasting the sustainability of Ireland's dairy herds, county by county, to 2036, and testing which policy levers would move it.
01 · The brief
From a paper to a map people can use
The paper, Spatio-Temporal Graph Neural Networks for Dairy Farm Sustainability Forecasting and Counterfactual Policy Analysis, treats Ireland's counties as a graph and forecasts how their dairy sustainability evolves over time.
AgriGraph implements it end to end, then carries the results to the people who would use them: an atlas to explore, a dashboard per county, benchmarks across years, and a way to test policy before applying it.
02 · Pipeline
Six stages, records in, decisions out
-
01
ICBF herd records
Herd records from the Irish Cattle Breeding Federation are the raw input.
-
02
VAE augmentation
A variational autoencoder augments the records.
-
03
4-pillar scoring
Each herd is scored on four operational pillars.
-
04 · core
Spatio-temporal GNN
Counties are graph nodes; the model forecasts each trajectory to 2036 with a 90% band.
-
05
Counterfactual levers
Policy levers re-run the forecast as what-if simulations.
-
06
FastAPI → React map
Results persist in PostgreSQL and feed an interactive atlas.
03 · What gets scored
Four pillars per herd
Every county dashboard plots them as a radar next to the forecast.
- Reproductive efficiency
- Genetic management
- Herd health
- Herd management
04 · The product
Three screens, one question each
Atlas
Where should I look?
Pick a county on the map to open its dashboard. The atlas ships with light and dark themes.
County dashboard
Where is this county heading?
Score and KPIs, the forecast trajectory to 2036 inside its 90% band, and the four-pillar radar.
Benchmarks
What would a policy change?
Slide across years against the national average and the highest and lowest county, initial vs. simulated policy.
05 · Stack
- Model
- Python
- PyTorch
- PyTorch Geometric
- Service
- FastAPI
- PostgreSQL
- Interface
- React
- Light + dark themes
