Notes on models
of the enterprise.
Essays and research notes from the Lyon team on foundation models trained on enterprise data, how they work, what they predict, and what they change about building ML inside a company.
-
Build vs. Buy: The Real Cost of Standing Up an Enterprise ML Team
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.
Read the essay → -
Churn Prediction with Foundation Models: Why Your XGBoost Ceiling Is Lower Than You Think
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.
Read the essay → -
Next-Event Prediction: How Transformers Learn the Grammar of Your Transactions
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.
Read the essay → -
LLMs Know the Internet. They Don't Know Your Business.
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.
Read the essay → -
What Is an Enterprise Foundation Model?
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.
Read the essay →