Transaction Foundation Model
Card networks and issuers run separate models for fraud scoring, next-merchant prediction, amount forecasting, dispute triage, and collections prioritization. Each has its own pipeline. This demo collapses them into a shared model with multiple surfaces. The default tab returns four signals from a single read of the customer's transaction history in around 22 milliseconds: fraud probability, next merchant, amount range, and MCC enrichment. The other tabs run alternate surfaces on the same model: dispute triage, collections prioritization, and fraud-pattern attribution. Adding chargeback likelihood or decline reason is a configuration change. Not a new model build. A stage-1 self-supervised pretraining step also lets the encoder catch roughly 5× more fraud on a newly onboarded entity with almost no label history — a synthetic-data demonstration shown in the Cold Start tab.
The Problem
Card networks and issuers maintain separate models for fraud scoring, next-merchant prediction, amount forecasting, and MCC enrichment. Each is its own pipeline, its own deployment, and its own inference cost. Discrete transaction features don't fit cleanly into a text-only LLM either.
How LFM Compares
Per-task gradient-boosted trees or transformer heads, each trained from scratch on tabular features. No shared representation across tasks. Every new business objective is a new model.
What LFM Unlocks
One model where today there are five or more. Fraud, dispute, collections, and fraud-pattern attribution share the same foundation, the same deployment, and the same audit trail. Adding chargeback likelihood, decline reason, or a new customer-segment signal is a configuration change. Every prediction also ships with a cited explanation suitable for an analyst note. Free-text analyst questions are on the roadmap.
Customer Archetype
This demo is fine-tuned on sample data. Results improve with your data.