Skip to content
AvraAvra
Esc
↑↓navigate↵open⌘Jpreview
On this page

Why Avra

Why decisions made on rows and snapshots miss what drives the outcome, and what a foundation model for relational data changes.

Your hardest decision has a graph underneath.

The thin-file company you can’t score. The fraud ring that passes every check. The customer deteriorating before your data says so. Each one shows up in the relationships around it, and not in the row your model sees.

The row is a summary

Most models are trained on one wide row per customer, account, or company: an aggregate of their tables, taken at one moment. Building that row throws away what the tables knew: who is connected to whom — owners, counterparties, suppliers, and the devices and addresses shared with other accounts — and when things happened, in what order, and how fast they’re changing.

The graph around C1C1O1O2K1same productT1support ticketD1shared deviceC9churnedflattenC1 as one rowcustomer_idC1orders_90d2avg_ticket182.40days_since_last12segmentgoldWho C1 is connected to, and when, is gone.
The same customer, as the graph around it and as one feature row. The row keeps a summary; the graph keeps who the customer is connected to, and when. Here, C1 shares a device with C9, which churned last month, and no column of the row says so.

It fails worst for the customers with little history. A company incorporated last week has an almost empty row, even when everything around it is known.

Every question restarts the project

A new question means a new dataset, new features, a new model, and a new evaluation. Little of that work carries over to the next question, so most of the questions a business could ask of its data never get a model.

Language models don’t close that gap. A prediction about a customer depends on the tables, keys, and timestamps that describe it; language models are built for text, and they don’t predict over a relational database.

The shift

Language got a foundation model. Then images, voice, and code. Relational data gets one now.

A foundation model is pre-trained once on broad data, then answers many different questions without a model built for each one. Avra RFM applies that idea to relational data. Pre-trained on diverse relational data, it reads your tables, keys, and timestamps as they are and predicts in context, with no training project in between.

What changes

More signal from the data you have. The relationships around a customer carry information its own row never will: who it transacts with, what changed nearby, how its neighborhood is moving. A model that reads that structure finds signal a feature table can’t express, including for customers with almost no history of their own.

Less engineering for each new question. The next question reuses what the model already knows: name the target, evaluate it on your history, and put the scores to work.

Context, not columns

Ownership, transactions, and shared infrastructure stay connected. Each customer is scored with its network around it, not as an average of it.

Time as input

A prediction sees your data as it was when the prediction is made, and nothing after. The model reads how a customer is changing, not one snapshot.

No file required

A company incorporated last week or an account opened this morning still sits inside a structure. The model reads that structure and returns a score.

The systems you run stay yours

No rip-and-replace, and no new workflow to adopt. Scores arrive in the credit, fraud, growth, and monitoring systems your team already operates, and your team keeps the controls: which version serves production, when to promote it, and when to roll back. Every model is backtested on your history and shadowed beside your current one before it decides anything. See Evaluation.

A frontier lab

Avra is an AI frontier lab. Its research teaches models to read relational data, and the work ships to production. Alongside the RFM, Avra runs a graph foundation model specialized for Brazil, Avra GFM, delivered through Avra Core. Read more at avra.ai/research.

Was this page helpful?