A relational foundation model, pre-trained on the temporal knowledge graph of a real economy, fine-tuned on contextual relational data, and made to generalize to entities it has never seen.
Avra's research is organized around a single bet: the right unit of intelligence for enterprise decisions is the relationship — modeled over time, learned once, and adapted inside each customer's workspace.
Avra's foundation layer is a relational foundation model — a graph neural network trained on a temporal knowledge graph of companies, individuals, counterparties, and events. It arrives pre-trained on a real economy, not as empty infrastructure waiting for your data. The objective is to learn structure — how decisions actually move — rather than to memorize features attached to single entities.
Tabular models stop where features stop. An inductive relational model reasons over the entities you do know to reach the entities you don't — businesses without a file, accounts before they incorporate, counterparties absent from the training set. Reach without retraining.
On top of the foundation, Avra fits customer-specific downstream models. The relational signal each enterprise sees in its own graph — payments, claims, applications, supplier links — conditions a workspace-exclusive set of weights. Data and weights stay isolated by tenant.
Decisions move through time. Avra's graph carries versioned edges and a temporal axis, so the model can ask what a relationship looked like before a default, before a fraud ring matured, or before a customer expanded. Trajectories carry signal that snapshots erase.
A growing list of research from Avra and the academic groups we collaborate with — on graph foundation models, inductive learning over relational data, and temporal reasoning at enterprise scale.