---
title: Avra GFM
description: Avra GFM reads every Brazilian company through the relationships around it, and returns probabilities of default and embeddings.
sidebar:
  label: Overview
---

No company stands alone. Each one sits inside a web of relationships: the people who own it, the other companies those people own, the sector it works in, where it operates, and the legal events around it. Those relationships often move before a company's own records do.

Avra GFM, the Graph Foundation Model, is Avra's foundation model of the Brazilian economy. It's pre-trained to read each company through the web around it, and through how that web changes over time. It turns what it reads into signals your decisions can use, probabilities of default and embeddings, delivered through [Avra Core](/core).

## A graph of the Brazilian economy

The GFM learns from a temporal knowledge graph that Avra builds and maintains. The graph connects companies to their owners and partners, to the other companies they're tied to through ownership, to their sectors and locations, and to the lawsuits they're part of. It also remembers: every relationship keeps its history, from the day it appears to the day it ends. Customer data never enters it.

<svg class="fig" viewBox="0 0 640 300" role="img" aria-label="A company at the center, connected to a person by ownership and control, to another company by ownership, to a location by geography, to a lawsuit by judicial events, and to an industry classification by sector. A second person is connected by a dashed line marked past ownership.">
  <line class="edge" x1="150" y1="56" x2="320" y2="150" />
  <line class="edge" x1="490" y1="56" x2="320" y2="150" />
  <line class="edge" x1="380" y1="150" x2="510" y2="150" />
  <line class="edge" x1="490" y1="244" x2="320" y2="150" />
  <line class="edge" x1="150" y1="244" x2="320" y2="150" />
  <line class="edge dash faint" x1="120" y1="150" x2="260" y2="150" />
  <text class="small label end" x="226" y="116">ownership and control</text>
  <text class="small label" x="414" y="116">ownership</text>
  <text class="small label mid" x="445" y="142">geography</text>
  <text class="small label" x="414" y="194">judicial events</text>
  <text class="small label end" x="226" y="194">sector</text>
  <text class="small label mid" x="190" y="142">past ownership</text>
  <rect class="soft" x="100" y="40" width="100" height="32" rx="16" />
  <text class="mid" x="150" y="60">Person</text>
  <rect class="soft" x="440" y="40" width="100" height="32" rx="16" />
  <text class="mid" x="490" y="60">Company</text>
  <rect class="soft" x="510" y="134" width="110" height="32" rx="16" />
  <text class="mid" x="565" y="154">Location</text>
  <rect class="soft" x="440" y="228" width="100" height="32" rx="16" />
  <text class="mid" x="490" y="248">Lawsuit</text>
  <rect class="soft" x="60" y="228" width="180" height="32" rx="16" />
  <text class="mid" x="150" y="248">Industry classification</text>
  <rect class="card dash faint" x="20" y="134" width="100" height="32" rx="16" />
  <text class="mid faint" x="70" y="154">Person</text>
  <rect class="solid" x="260" y="134" width="120" height="32" rx="16" />
  <text class="title mid on-solid" x="320" y="154">Company</text>
  <text class="small label mid" x="320" y="290">Dashed: a relationship that ended, kept in the graph's history</text>
</svg>

The kinds of entity in the knowledge graph around one company, and the relationships that connect them. A relationship that ended, such as a past owner's, stays in the graph's history.

## Context no single record holds

A company's own records describe what it has done. The graph shows what surrounds it: the companies it shares owners with, the lawsuits around them, its sector, and its region. The GFM reads that context for every company, so even a company founded last month arrives with signal: it already has owners, a sector, and a location.

Because the graph keeps its history, the GFM can also look at a company as it stood on any past date. That's what makes a backtest on your own history honest, because each prediction sees the graph as it was then.

## What it delivers

**[Probabilities of default](/use-cases/credit-intelligence)**

For every company, its probability of default at five horizons, from 30 days to a year, for policy, pricing, and portfolio monitoring.

**[Embeddings](/gfm/embeddings)**

Each company's place in the graph as a vector, ready to add to the models you already run for credit, fraud, and acquisition.

## Make it yours

Your portfolio has its own definition of default. Avra builds a credit model for it on GFM embeddings together with your features and outcomes, so its probabilities describe your events, and it can add a 0–1000 score for your decision rules. See [Credit](/use-cases/credit-intelligence).

## Put it to work

Send a company's CNPJ and get its outputs back: one company at a time through the [API](/core/predictions) when a decision comes up, or a whole portfolio in [batch](/core/batch-inference) on your schedule. The GFM runs in Avra's cloud, and new versions arrive as governed releases that you adopt when you're ready. See [Model versioning](/core/model-versioning).

```mermaid
flowchart LR
    C["A company's<br/>CNPJ"] --> G["Avra GFM"]
    G --> P["Probabilities<br/>of default"]
    G --> E["Embeddings"]
    P --> D["Your credit<br/>decisions"]
    E --> M["The models<br/>you already run"]
    classDef accent fill:#0A01FE,stroke:#0A01FE,color:#fff
    class G accent
```

## GFM and RFM

Avra GFM starts from a company's CNPJ and reads Avra's knowledge graph around it. [Avra RFM](/rfm) starts from your own tables and predicts the outcomes you define over them. See [Models](/models) for how they compare and which fits your question.

## Where to go next

**[Credit](/use-cases/credit-intelligence)**

Read and use probabilities of default at five horizons.

**[Build with embeddings](/gfm/embeddings)**

Add GFM embeddings to the models you already run.

**[Live predictions](/core/predictions)**

Call the GFM with a company's CNPJ.

**[Avra RFM](/rfm)**

Predict any outcome from your own relational data.
