Which advantages over its rivals hold up, and why are they hard to copy?
Lyon's models hold an advantage in richer signal: they learn how events relate to one another over time in ways that manually engineered features can miss, and at a consumer fintech with tens of millions of users a Lyon model produced a 112% relative improvement in default-prediction PR-AUC over the incumbent tree-based model.
Lyon applies the transformer technology behind large language models to transaction histories, loan records, and customer activity. Its downstream models can also take existing engineered features alongside the learned embeddings.
Lyon trains and serves models inside the customer's own cloud, and the customer keeps its data and every model trained on it.
Which risks could break the thesis, and what would warn of them?
Financial firms building their own transaction-data models could impair Lyon's growth, because each in-house program removes an institution that would otherwise need an outside model provider. Mastercard started developing its own LTMs trained on billions of anonymized transactions, and a wave of financial firms are creating bespoke gen-AI models specifically for transaction data.
Lyon builds private foundation models for banks, fintechs, funds, brokers and payments companies, turning their own data into better predictions. Training and inference run inside the customer's VPC, and the customer's data stays with the customer, along with every model trained on it.
How does it make money per customer, and what is still unknown?
A single trained model serves several product lines, answering churn, credit risk, fraud, collections, income and cross-sell in place of separate feature-engineering projects, so one deployment creates value beyond the single task on which it was measured. At a consumer fintech with tens of millions of users, that model improved default-prediction PR-AUC by 112% relative to the incumbent tree-based model.
Each model is trained and run inside the customer's own VPC, and the customer retains the data and every model built on it. Lyon develops and deploys each institution's model and supplies software that keeps the model current and integrates its predictions.
What does the product do for its buyer, and how is it delivered?
Lyon builds a private foundation model tailored to each bank, fintech, fund, broker and payments company, training it from scratch on that institution's event data, hosting it in the institution's own cloud, and transferring ownership to the institution.
Lyon develops and deploys each institution's model alongside software that keeps the model current and integrates its predictions into the institution's existing systems.

Lyon's foundation models for banks, fintechs, funds, brokers and payments companies run on own event data in the customer's cloud.