Lyon

Companies that want to predict churn, credit risk, fraud, collections, income, or cross-sell today typically build a separate model for each task, each requiring its own months-long feature engineering project. This duplicative effort slows down teams that need multiple predictions from the same underlying customer data. Sending sensitive customer event data to third-party vendors for model training also raises data privacy and ownership concerns for these companies.

Lyon trains a foundation model from scratch on a customer's own event and profile data, producing customer embeddings that serve as a shared input for many downstream prediction tasks, including churn, credit risk, fraud, collections, income, and cross-sell. Teams that already have working features can combine them alongside the embeddings rather than discarding them. To adapt the model to a specific task, users select a label table, verify the data split, and run a LoRA fine-tune to train a task-specific adapter. Training and inference both run inside the customer's own VPC on AWS, GCP, or Azure, so the underlying data never leaves the customer's cloud, with the company citing SOC 2, GDPR, and LGPD compliance.