What is a graph foundation model?
A model pretrained to predict over relational, connected data, so it can answer new predictive questions on tables it has never seen, without task-specific training.
Resources
Learn the ideas behind graph AI.
Short, plain explainers of the concepts that make production-grade graph AI work. No jargon for its own sake.
Start with the fundamentals.
How relational data becomes a graph
Rows are entities, foreign keys are edges, and timestamps order events. Your existing schema already describes a temporal graph; nothing new to model by hand.
The Predictive Query Language
PQL expresses a prediction as a single statement: what to predict, for which entity, over what horizon. The engine handles features, training, and scoring.
Point-in-time safety
Anchoring every example to a cutoff means the model only sees what was knowable then. It is what separates an honest backtest from a leaky one.
Graph AI vs. traditional ML
Hand-built pipelines flatten relationships into features and lose multi-hop signal. Graph AI learns across tables directly, so the signal between rows survives.
Backtesting, honestly
A held-out temporal split, reported per prong, tells you what would really have happened. A training-set score does not.
Ready to try it on your data?
Bring a labelled table or use a bundled sample, and we will run a use case end to end.
We will reach out within one business day.