Avra at Bloomberg Línea SummitNov 23

Avra predicts business outcomes from your connected data. One foundation model, ready for any question. No model to build for each one.

Merchant 1043 · 0.81linked · Account 2207

A relational graph drawn as columns of bars: records are clusters, relations are trails, flowing into the card where the answer arrives. The card is an interactive example on sample data: type a question about churn, default, fraud or conversion, or pick one of four examples, and Avra RFM ranks the records with a prediction and says why the first one ranks first: something that happened to a record linked to it.

Interactive exampleSample data
  1. Merchant 1043High0.81WhyLinked Account 2207 closed last week
  2. Merchant 2210High0.67
  3. Merchant 0877Medium0.54

Leading banks, telcos and marketplaces already use Avra, where a decision turns on who is connected to whom.

−70% SME card defaults, with95% approval kept. Read the case

The shift

The shift: relational data gets its foundation model.

Most of what a business knows sits in tables linked by keys: customers, accounts, transactions and the time between them. Avra RFM reads your database as it is, so it starts from structure and predicts with no model to build first.

Read the research

How it works

From your tables to a decision, in three steps.

Avra RFM is pre-trained on diverse relational data, so there's no model to build in between: connect, ask your question, and put the predictions to work.

  1. 01

    Connect your tables

    Point Avra at your database as it is: every table, every key. No feature engineering, no flattening into one wide row.

  2. 02

    Ask the question

    Write the question you want answered, then evaluate it on your own history before anything goes live.

  3. 03

    Act on the predictions

    Every customer, account or company ranked, each with an explanation, delivered by API or batch to the systems you already run.

Documentation

The platform

It runs inside your systems.

Live or in batch, as predictions or as features, proven before it decides, and put to work on a canvas your team runs.

Live predictions

A prediction for every application, as it arrives.

Your decision engine sends a request and gets back a prediction, the relationships behind it and the model version that made it.

Real time, with reasons

  • Per-request API
  • Reasons on every prediction
  • Pinned model versions
  • Any entity in your schema
Learn more

Batch predictions

Rank every account before the day starts.

On the schedule you set, Avra predicts for every customer or account and writes the results to your warehouse, beside the tables they came from.

On your schedule

  • Daily, weekly or on demand
  • Written to your warehouse
  • Point-in-time cutoffs
  • No feature pipeline
Learn more

Embeddings

Add the relationships to the models you have.

Put Avra's relational embeddings beside your own features, and your credit, churn or pricing models see who is connected to whom.

Into your own models

  • Late fusion
  • Lookalikes by structure
  • Any downstream model
  • One vector per entity
Learn more

Governance

Prove it before it decides anything.

Every model is backtested on your history and shadowed beside your current one before it decides anything, then held to the band you set.

Governed by default

  • Backtests on your history
  • Shadow deployment
  • Production monitoring
  • Bands you set
Learn more

Design the decision

Turn the prediction into the decision.

Describe the workflow in a sentence and Studio drafts it on a canvas: your policy, your review queue, your systems, with an explanation behind every outcome.

Avra Studio

  • Drafted from plain words
  • Set off by events
  • Your policy as steps
  • Reasons in every outcome
Learn more

Your customer, in five places.

Systems that never meet, and a model to build for every question you ask of them.

An example on sample data. Five tables from five systems, customers, accounts, transactions, support tickets and app devices, lie apart, with the work it takes today to answer one question: a join that has run for hours, an exported features file, a training notebook, and a new question that becomes another project. As the page scrolls, Avra RFM joins the tables by the keys they share and follows customer 1043 across them. Then it adds a churn prediction for every customer as a new column: 0.81 for customer 1043, because account 2207, linked to them, closed last week, they asked support how to close their account, and their card has not been used since October 6.

Your hardest decision has a graph underneath.

Talk to sales

Ask about Avra on

Merchant 1043