Models
How Avra's two foundation models compare — Avra RFM for your own relational data, Avra GFM for Brazilian companies — and which fits.
Avra has two foundation models. Avra RFM reads your own relational database and predicts the outcomes you define. Avra GFM reads Avra’s knowledge graph of the Brazilian economy and returns signals for any Brazilian company. Both reach your systems through Avra Core, live or in batch, and Avra RFM can also run inside your AWS account.
How they compare
| Avra RFM | Avra GFM | |
|---|---|---|
| What it is | The foundation model for relational data | A graph-based foundation model, specialized for Brazil |
| Reads | Your relational database, any schema | Avra’s knowledge graph of the Brazilian economy |
| Starts from | Your tables, keys, and history | A company’s CNPJ |
| You receive | A score for each record, for a yes-or-no outcome, or an estimate, for a number | Probabilities of default at five horizons, from 30 days to a year, and a 1024-dimensional embedding |
| Specialized for you | Reads your history in context by default, and a copy can be adapted or fine-tuned for your workspace | A credit model built for your portfolio on its embeddings, your features, and your outcomes |
| In Avra’s cloud | Through Avra Core, live or in batch | Through Avra Core, live or in batch |
| In your AWS account | On Amazon SageMaker, through AWS Marketplace | Runs in Avra’s cloud only |
Which model fits
Start from where the answer lives: in your own records, or in the context around a company.
| You want to | Use |
|---|---|
| Predict an outcome you define, such as churn, delinquency, conversion, or fraud, for the records in your own tables | Avra RFM |
| Estimate a number from your own history, such as how much each customer will spend next month | Avra RFM |
| Know a Brazilian company’s probability of default, at horizons from 30 days to a year | GFM probabilities of default |
| Give the models you already run the relationships around each company | GFM embeddings |
| Learn your own definition of default from the GFM’s view of each company | A credit model on GFM embeddings |
The two work side by side. A lender can predict delinquency over its own loan book with Avra RFM, and read Avra GFM’s probabilities of default for the companies applying, both through the same API.
How you reach them
Both models reach your systems through Avra Core: one record at a time through the live API when a decision comes up, or a whole list in batch on your schedule. Each model has versions, and an alias such as production decides which one answers, so your systems move to a new version only when the alias does. See Model versioning.
Avra RFM can also run inside your AWS account, on Amazon SageMaker. See Avra RFM on AWS SageMaker.