Skip to content

Vertical foundation models / Payments

  • Input → outputTransactionsto4 signals
  • Model350M backbone + encoder
  • Runs onCloud GPU
  • What it doesFraud · merchant · amount · category
FraudNext merchantAmountMCCOne frozen backbone · four task heads

One backbone.Many payment signals.

Read transaction history once. Predict what matters next.

Reference demo on synthetic data. Explore the payment tasks below.

Transaction historyLiquid Payments modelFraud · Merchant · Amount · MCC
Four signals from one read of the customer, in around 22 milliseconds. Other tabs above run alternate surfaces (dispute, collections, fraud pattern) on the same model.

Customer Archetype

Loading archetypes…
Select an archetype to run multi-head inference.

Problem

A payment is part of a pattern. Card-testing bursts, changing merchants, and unusual spending emerge across transactions. Separate models for fraud, merchant prediction, amount forecasting, and category enrichment each need their own training and maintenance. The same customer history gets processed again for every task.

Solution

A structured encoder connects transaction history to an LFM backbone. The multi-head demo returns fraud probability, next merchant, amount range, and merchant category together. Dispute, collections, and fraud-pattern models reuse the same backbone architecture with task-specific adapters and heads. New payment tasks can build on that foundation instead of starting from scratch.

01

See the sequence.

Read relationships across 64 transactions, including timing, merchants, amounts, and device changes.

02

Share the foundation.

Adapt a frozen Liquid backbone with lightweight adapters and task heads for different payment workflows.

03

Start with fewer labels.

Explore the synthetic cold-start study to see how self-supervised pretraining helps when labeled history is scarce.

Start building with Liquid

Model catalogContact usTry our API