Pular para o conteúdo
AvraAvra
Esc
↑↓navegar↵abrir⌘Jpré-visualizar
Nesta página

Welcome to Avra

Avra builds the foundation model for relational data, the linked tables every business keeps, and puts its predictions into the decisions you already make.

Ask your database what happens next

Avra is an AI frontier lab building the foundation model for relational data. Avra RFM, the Relational Foundation Model, reads your database as it is and predicts what happens next — which applicants will default, which merchants will churn, which leads will convert — every question from one pre-trained model, without a training project for each.

What relational data is

Every business runs on relational data. It’s how the systems you already use store what you know: in tables, where each table holds one kind of thing and each row describes one of them. A customers table has a row for each customer, an accounts table a row for each account, and a payments table a row for each payment.

Keys link the tables. Every row has an identifier of its own, and rows point at each other by carrying those identifiers: a payment carries the ID of the account it came from, and the account carries the ID of the customer who owns it. Most rows also record when they happened, or when they were true.

customerscustomer_idsinceC12021-03-14C22023-11-02accountsaccount_idcustomer_idA1C1A2C1A3C2paymentspayment_idaccount_idpaid_atP1A1Apr 02P2A1May 02P3A3May 09keysC1C2A1A2A3P1P2P3customeraccountpayment
The same rows, as tables and as a graph. Every row becomes a node, and every key that points at another row becomes an edge.

Follow the keys and the tables become a graph: every row is a node, and every key that points at another row is an edge. Read with the timestamps, that graph tells you who bought what, who paid whom, who owns which company, and in what order it all happened. See Relational data.

The answer to most business questions sits in those links. Whether a company defaults depends on its own payments, and also on its owners’ other companies, its largest customers, and its suppliers. Whether a new account is fraud depends on the devices and addresses it shares with accounts you’ve already seen.

Most models never see the links. They’re trained on one wide row per customer, account, or company: a summary of its tables, taken at one moment and built by hand for one question. The relationships and the order of events are flattened away, and the next question starts the work again. Avra RFM reads your tables, keys, and timestamps as they are, with no features to build first. See Why Avra.

How your data reaches Avra

You send the tables you already keep, as they are: the records you want to score, the tables related to them, the keys that link them, the columns that say when each row was true, and your historical outcomes, such as defaults, conversions, or confirmed fraud. Upload them as files, or have them read from the storage where they already live.

Every delivery goes through Avra’s ingestion pipeline. It checks each one against the data contract agreed with your Avra team and publishes a new version only when it passes, so a model only ever sees data that passed. It also keeps when each row was true and when it was recorded, so your data can be replayed as it stood on any past date, and a backtest sees only what was known then. See Time and history.

The pipeline runs in Avra’s cloud, inside your isolated workspace, or in your own environment, next to your data. Your data is protected in both, and nothing about your business trains another customer’s model. See Deployment.

What you get back

For each record you ask about — each customer, account, or company — Avra RFM returns a score that ranks it against the others. Avra Core delivers the scores to the systems you already run: one record at a time through the API when a decision comes up, or a whole portfolio in batch on your schedule. Your team keeps the controls: which version serves production, when to promote a new one, and when to roll back.

Avra also runs a second model for Brazilian companies: Avra GFM, the Graph Foundation Model, pre-trained on Avra’s knowledge graph of the Brazilian economy. Given a company’s CNPJ, it returns embeddings and probabilities of default through Avra Core. See Models for how the two compare.

Where to go next

Building an integration? Go to the API reference.

Esta página foi útil?