In finance, marketing, and sport, the richest signals aren't in text — they're buried in transaction histories, event sequences, ownership graphs, and order flow. Foundation models learn latent features from that data no analyst could ever curate. A single model that spans fraud detection, marketing, and operations — with uplifts that can be dramatic. Transparent enough for the regulator.
A foundation model is a neural network trained with self-supervision: rather than learning from labelled examples, it finds patterns within the data itself — patterns hidden to traditional data science approaches. The same approach applies wherever you have large volumes of structured behaviour — tables, graphs, and sequences. Most business data is one of the three. Language models are just a subset of foundation models, not the definition of them.
Trained on hundreds of billions of transactions across its network, Stripe built a single foundation model that generates a representation of each payment, then fine-tunes that representation for fraud detection, authorization optimization, and risk scoring — replacing a collection of narrower, hand-tuned models.
Nubank trained transformer-based models on customers' transaction sequences, then extended them into the graph. By propagating signals across the network of who transacts with whom, the model builds rich representations for new or thin-file customers from their neighbours — enabling earlier and more precise credit decisions than traditional scorecards.
Language models assume meaning lives in a stream of text. Most of the data companies sit on doesn't look like that. We work with three architecture families — each matched to how the data is actually shaped, not forced into a format it wasn't designed for.
Ownership structures, beneficial owners, director networks, supply chains, entity relationships. We pretrain node and edge representations across the full graph, then fine-tune for a specific task — link prediction, anomaly detection, shell-company and obfuscation detection — without retraining from scratch each time the task changes.
Order books, bet placement, market events, transaction streams. We treat each entity's history as a sequence — the way a language model treats a sentence — and pretrain on the structure of what tends to follow what, then fine-tune for forecasting, anomaly and manipulation detection, or signal extraction.
Spreadsheets, database tables, ledgers. We use TabPFN — a foundation model pretrained across millions of real-world tables by Prior-Data Fitted Networks — as our starting point, then fine-tune it on your data for the specific task at hand. Unlike our graph and sequential work, where we train from scratch on your proprietary data, TabPFN's pretrained weights already encode deep cross-column reasoning that gradient-boosted models have to be hand-engineered to approximate. The result is strong performance with minimal data and no manual feature prep.
These are working case studies, not finished products — built to test where each architecture earns its keep against a strong, boring baseline before we'd ever recommend it to a client.
Our graph foundation model, pretrained on the UK Companies House ownership network, identifies companies likely to dissolve with a 17.5% lift over a well-tuned gradient-boosted baseline — using structural signals across director, shareholder, and PSC relationships that tabular models simply can't access. The same pretrained graph can be fine-tuned on proprietary credit, churn, or underwriting data to predict the business outcomes that matter to your team, without starting from scratch.
Our sequential foundation model, trained on Polymarket order flow, learns the structure of how participants behave over time — not just what they trade, but when, how, and alongside whom. Fine-tuned heads then categorise funders by behavioural profile, flag accounts exhibiting patterns consistent with insider trading, and predict market outcomes ahead of resolution. The same pretrained model can be adapted to proprietary trading or event data to surface the same signals in closed markets.
A foundation model is only worth deploying if it beats what you already have — including a well-tuned gradient-boosted baseline, which is a higher bar than most vendors admit.
We benchmark your current model or rules engine against a foundation-model approach on your own data, over a short fixed-scope pilot, before any commitment to build.
If the lift is real, we pretrain a model matched to your data's actual shape — graph or sequential — and fine-tune it for the specific task that matters to your business.
You keep the trained model, the evaluation harness, and the documentation needed for your own team to maintain and extend it going forward.
If you're sitting on a graph, a transaction history, or an event log that your current models aren't doing justice to, we'd like to hear about it.