Vertical Foundation Models
One frozen 350M backbone per vertical. A small encoder teaches it your data, and every business objective is an adapter, not a rebuild.
talk trackFraud, next-merchant, amount, and MCC usually mean four separate models. Here all four read one forward pass of the same frozen backbone.
One forward pass · every head reads the same backbone state · a new business objective is a new head, not a new model
Four payment signals, one forward pass
One transaction encoder turns 64 transactions × 15 features into 960 pseudo-tokens the frozen backbone processes with attention. Four heads predict in parallel: fraud, next-merchant, amount bucket, and merchant category. Adding chargeback likelihood or decline-reason is a head, not a model build.
Liquid team only. These instances run on our GPU and an SA drives the session — ask your Liquid contact and they will run it with you.
- 64 tx × 15 features → 960 pseudo-tokens (d=1024)
- 4 heads: fraud · next-merchant · amount · MCC
- ~16M trainable on a 350M frozen base
- ~22ms warm on the demo H100
Working demonstrations of the recipe on synthetic or public data (Medicare vertical: CMS DE-SynPUF, deidentified) — demo-grade previews, not shipped products or production benchmarks.