https://lyon.so/blog/ Essays and research notes from the Lyon team on enterprise foundation models: training transformers on transactional data, prediction, and applied ML strategy. en-us Wed, 01 Jul 2026 09:00:00 -0700 https://lyon.so/blog/build-vs-buy-enterprise-ml/ https://lyon.so/blog/build-vs-buy-enterprise-ml/ Wed, 01 Jul 2026 09:00:00 -0700 The sticker price of an in-house ML team is salaries. The real price is the two years your best engineers spend on data plumbing instead of your product, and the second model, which costs almost as much as the first. https://lyon.so/blog/churn-prediction-foundation-models/ https://lyon.so/blog/churn-prediction-foundation-models/ Tue, 23 Jun 2026 09:00:00 -0700 Gradient-boosted trees on hand-built features hit a ceiling that no amount of feature engineering breaks. A model pretrained on the full event history sees churn coming weeks earlier, because it never threw the sequence away. https://lyon.so/blog/transformers-on-transactional-data/ https://lyon.so/blog/transformers-on-transactional-data/ Fri, 12 Jun 2026 09:00:00 -0700 A technical walkthrough of how enterprise events become tokens, why next-event prediction is the right self-supervised objective for behavioral data, and what the model has to learn to get good at it. https://lyon.so/blog/llms-dont-know-your-business/ https://lyon.so/blog/llms-dont-know-your-business/ Thu, 04 Jun 2026 09:00:00 -0700 Language models are trained on public text, and the signal that runs an enterprise isn't text. Why RAG and agents don't close the gap, and what a model that actually knows your business looks like. https://lyon.so/blog/what-is-an-enterprise-foundation-model/ https://lyon.so/blog/what-is-an-enterprise-foundation-model/ Tue, 26 May 2026 09:00:00 -0700 A plain-language definition of the category: one model, pretrained on a single company's own event data, that every downstream prediction fine-tunes from. What it is, how it's built, and why it exists now.