Skip to main content

Foundation models

Closed beta. Weights and inference access are granted per engagement. Reach out to scope a pilot.

Language models are foundation models of text. We are building the equivalent for behavior: models pretrained on large volumes of real human interaction sequences, which transfer to a new product or catalog with little task-specific data.

The bet is the same one that worked for language. A model trained broadly enough on how people act stops needing a bespoke architecture per application, and a downstream team gets a strong starting point on day one instead of after a year of collecting their own labels.

First release: recommendation

Recommendation is where behavioral data is densest and where the evaluation is least ambiguous, so it is where we started.

The model is a sequence model over user interaction histories. It consumes a user's ordered actions and predicts what comes next, in the same autoregressive way a language model predicts the next token. Items are represented as semantic IDs rather than as rows in an embedding table, which is what lets the model transfer across catalogs at all.

PropertyWhat it buys you
Cold startA new item is representable the moment it has content. No interaction history required before it can be recommended.
TransferPretrained on interaction data across domains, then adapted to your catalog. Useful before you have accumulated your own signal.
Generative retrievalThe model generates an item identifier directly. Constrained beam search over the codebook replaces approximate nearest neighbor over an index.
Scales with catalogParameter count is set by the codebook, not by the number of items. A catalog can grow without the model growing with it.

How you would use it

  1. Tokenize your catalog. Fit a semantic ID tokenizer on your item content. See Semantic IDs.
  2. Adapt on your sequences. Supervised fine-tuning on your user histories, expressed as next-token prediction over semantic IDs.
  3. Align to your objective. A reinforcement learning pass against your actual ranking objective, not a proxy. What counts as a win is your definition, set at onboarding.
  4. Serve. Call it through the API, or take the weights and serve them yourself. See Models for the catalog and jean-rec-1 for the first release.

You can skip straight to step four. The pretrained backbone answers /predict and /embed with no tokenizer and no adaptation, which is the cheapest way to find out whether any of this works on your data.

Beyond recommendation

Recommendation is the first head, not the thesis. The same pretrained behavioral backbone is what the other downstream tasks run on: matching, personalization, memory, and simulation of populations.

We are deliberately not shipping those as products yet. They come after the backbone is good enough that they are adaptations rather than separate systems.

Status and access

StageClosed beta
AvailableRecommendation head, per engagement
DeploymentHosted API, or weights in your environment
PrerequisiteA structured interaction dataset. We can help build one, see Data.

Request beta access: jonathan@jeantechnologies.com