Overview
One model for every question in your database. Credit, fraud, growth, and retention are applications, not separate products.
Credit, fraud, acquisition, and retention are applications of the same models, not separate products to adopt. Each use case starts from a decision your business already makes, turns it into a prediction problem, and puts the scores back into the system that makes the decision. If you can name the records to score, the outcome, and the window it must happen in, Avra RFM can predict it from your tables.
From a question to a prediction
Every use case in this section starts from a business question and turns it into a prediction problem with three parts: who is scored, the outcome, and the window the outcome must happen in. The map traces each question to its problem and to the page that covers it.
| Stage | The question | Start here |
|---|---|---|
| Acquire | Which leads convert this month? | Lead scoring |
| Acquire | Which audiences bring customers worth keeping? | Media acquisition |
| Acquire | Which field visits close a deal? | Field sales |
| Lend | Which applicants default within 90 days? | Credit |
| Protect | Which new accounts belong to a fraud ring? | Fraud detection |
| Retain | Which customers churn in the next 28 days? | Talk to us |
| Any | Your own target, in the models you already run | Build with embeddings |
To turn a question of your own into a target, see What you can predict.
How every use case works
Name the decision and the target
Start from a decision your business already makes, and write it as a prediction problem: the records to score, the outcome that matters, and the window it must happen in. Defaulted within 90 days, converted within 30, churned within 28: your historical outcomes define the target.
Bring the tables
Send the tables you already keep, linked by their keys, with the columns that say when each row was true and when it was recorded, and your past outcomes as labels. Every delivery goes through data ingestion, which checks it against your data contract. See Relational data.
Backtest on your history
The model predicts at past dates whose outcomes are known, seeing only what was known on each date, and its results are compared with the model you run today. See Evaluation.
Shadow, then promote
The model runs beside your current one on live traffic, with no effect on decisions. When the comparison is settled, reassign the production alias to the version that earned it. Rolling back is reassigning it again. See Model versioning.
Put the scores to work
Score one record at a time through the API when a decision comes up, or a whole list in batch on your schedule, and let your policy turn scores into actions.
How scores reach a decision
Avra returns scores, and your systems make the decision. The tables and outcomes that give a prediction its context arrive through data ingestion. The record to score goes straight to the model in an API request or a batch file. Your policy turns the score into an action, and the outcomes you record afterward become labelled examples for later predictions once their window has closed.
Each use-case page follows that flow in the same order: the decision and why relationships matter for it, an illustrative schema of the data behind it, the prediction problem, how scores reach the decision, how the model is evaluated, and how to integrate it.
Two ways in
| Avra RFM on your data | Avra GFM outputs | |
|---|---|---|
| Reads | Your tables, any schema | Avra’s knowledge graph of the Brazilian economy |
| Who it scores | The records in your tables: applicants, accounts, leads | Brazilian companies, by CNPJ |
| Returns | A score for your target | Embeddings and probabilities of default (PDs) |
| Specializing it | Optional adaptation or fine-tuning of a copy exclusive to your workspace | A separate downstream model trained on GFM embeddings with your features and outcomes |
| Runs in | Avra’s cloud, or your AWS account | Avra’s cloud only |
The first route works for any target you can name. Avra RFM predicts over your own tables in context, and where it pays, a copy is adapted or fine-tuned for your workspace. The second serves Brazilian companies: Avra GFM supplies embeddings and PDs through Avra Core, and a separate downstream model can combine its embeddings with your features and outcomes. See Models for how they compare.
Live, batch, or embeddings
| Path | Use it when | For example |
|---|---|---|
| Live predictions | A decision comes up one record at a time | Underwriting an application, screening a new account, scoring a signup |
| Batch inference | You score a whole list on a schedule | Rescoring a portfolio overnight, ranking a sales territory, enriching a marketing list |
| Embeddings | Your team runs its own models | Adding GFM embeddings to a credit, fraud, or lookalike model |
Where to go next
Lead scoring
Rank leads by the structure your top customers share, and route them by score.
Media acquisition
Send your ad platforms the signal they can’t see on their own.
Field sales
Put your reps on the accounts most likely to buy.
Credit
Default risk across five horizons, including companies with little history of their own.
Fraud detection
Rings and collusion that a single row can’t show, ranked into the queue your analysts work.
Build with embeddings
Add Avra’s embeddings to the features your own models already read.
Contact your Avra representative to scope a target that isn’t listed here. For the terms used across these pages, see the glossary.