AI Frontier Lab

Graph Foundation Modelfor the economy.

Avra builds predictive models for enterprise decisions. We see the economy as it is — not rows in a table, but nodes in a network. The same foundation powers credit, fraud, growth, and any decision that depends on understanding the relationships between them.

Book a demoOur Thesis
Backed by frontier funds. Shipping with enterprise design partners.
Scout
Scout
Case studies marked — click through to read the brief
Inductive reach

Reach the entities others can't see.

Tabular models stop where features stop. A relational model reaches further — by reasoning over the entities you do know to reach the entities you don't. Companies before they incorporate, accounts without a file, businesses the bureau can't see.

About the foundation
01

No-file scoring

An entity with zero direct features is still scored — through its counterparties, its position in the graph, and the behaviour of the accounts connected to it.

02

Under-modeled businesses

Small and mid-market businesses sit in the gap a bureau can't reach. The relational signal already moves through their accounts. Avra reads it where others stop.

03

Pre-incorporation entities

Score a company before it is legally formed — before a CNPJ exists. The model reasons over founder graphs, supplier links, and counterparty history, not just a tax ID.

04

Inductive generalization

The model generalizes to entities it has never seen. The relational structure transfers — you don't have to be in the training set to be scored.

Reason over the entities you do know to reach the ones you don't
Brazil-native foundation

Brazil's economy, modeled as a temporal graph.

Avra pre-trains on public and licensed signals — companies, owners, counterparties, judicial events, obligations, geography — then fine-tunes that foundation inside each workspace. The moat is not a larger table. It is the structure underneath the decisions.

01

Country-scale coverage

Public and licensed signals across active companies, owners, counterparties, places, and events — connected as one temporal economy.

02

Institutional relationships

Ownership, control, supplier paths, judicial and regulatory events, obligations, and geography are modeled as graph structure, not appended as flat enrichment.

03

Time-aware by default

Relationships are evaluated as they existed at the decision date. Backtests and production scores do not learn from the future.

04

Brazilian semantics

CNPJs, branches, corporate groups, marketplaces, informal networks, and legal events carry local meaning. The foundation model starts there.

05

Workspace adaptation

Customer data stays isolated. The Brazil-trained foundation is adapted into customer-specific models with exclusive downstream weights.

One foundation. Fine-tuned for every decision your team makes.

Credit

Avra's score finds the good customers a traditional bureau evaluation leaves out. The pilot generated 1.8× NII on the Avra-scored cohort against control on the same book.

1.8×net interest income vs. control

Solution
Credit · model

Each iteration absorbed new risk signal from the relational graph. The final relational model beat the internal model by +2.5 p.p. consolidated — and +5.4 p.p. on the held-out test set.

+5.4 p.p.AUC over the internal model, on test

Solution
Fraud

Avra embeddings correlate just 18–22% with the hundreds of features the customer already has in production, capturing risk dimensions proprietary data can't see on its own.

+5 p.p.ROC AUC, orthogonal to internal features

Solution
Growth · acquisition

Avra's propensity score selects the CNPJs most likely to activate. The campaign directed at that cohort nearly doubled conversion against baseline — with no new tracking pixel to install.

+90%conversion lift on Avra-prioritized cohorts

Solution
Growth · field sales

Avra ranks the pipeline by propensity and lead quality. Field sales narrowed effort to the best leads and converted at 18% — versus an 8% baseline on the broader book.

2.25×conversions with 8× fewer touches

Solution

Developer-first, enterprise-ready.

Integrate Avra into your decision path in hours, not weeks. API, SDK, and playground — with the audit trail and tenant isolation enterprise teams require.

Read the docs →
  • AAPI

    Predict and explain endpoints. Documented, versioned, and tenant-scoped.

  • PPlayground

    Experiment with real decision flows in your browser. Replay traffic, inspect evidence, fork model surfaces.

  • EEnterprise
    LGPD compliantISO 27001 in progressTenant-isolatedAudit-ready
Deployment
Managed
YOUR ENVIRONMENTAVRA CLOUDYOUR APPAPI · BATCHControl planeBILLINGDASHBOARDWEBHOOKSData planeFEATURE STOREINFERENCETRAININGmTLSmTLS
Data planeControl plane

Both the control plane (billing, dashboard, webhooks) and the data plane (feature store, inference, training) run inside Avra Cloud. Your application calls a managed endpoint over mTLS.

From the founder

The economy is not tabular.

Companies are not rows in a table — they are nodes in a network, and that network is the signal.

—  Bruno Alano ·  Co-founder & CTO ·  ex-OpenAI
Talk to us

Bring us a decision that should be smarter.

We work with a small group of enterprises running decision systems at real scale. If credit, fraud, or growth is the difference between a good quarter and a great one, we should talk.

Book a demo