Your model · Your data · Your cloud
Lyon trains a foundation model on your company's own data. A single model that learns how your business actually works: who your customers are, where risk and opportunity concentrate, and what happens next.
For decades, enterprise prediction has meant hand-built features, heuristics, and tree-based models: analysts decide what matters, and the model sees only what they chose. Vision, speech, and language all learned the same lesson. Give a foundation model enough data and compute, and it beats hand-crafted knowledge every time. Your business already has the data: transactions, orders, claims, sessions, payments, millions of events.
Lyon pretrains a foundation model on that data, using the same transformer architecture behind LLMs. The result is a model of your company that has read every event in your history and learned the grammar behind it. From that single model, every predictive question your teams ask becomes a fine-tune from that foundation.
Will this customer churn? Which claim is fraud? What sells next quarter? One model answers all of it.
A foundation model is trained on your data: every transaction, order, claim, and interaction, every customer, every year of history. Along the way it comes to understand customer segments, seasonality, risk, and intent. What comes out is a rich, detailed embedding for every customer and account in your business.
Churn, credit risk, fraud, lifetime value, propensity, segmentation: they all come from the same model. Each one is a downstream model fine-tuned from those embeddings, so a new prediction is a quick adaptation, not a new project. Already running your own models? Combining our embeddings with your existing features is always an option.
A single model of every account, pretrained on the full transaction history of every customer.
Orders, sessions, returns, and campaigns become one behavioral sequence per customer.
Risk priced from the full history of policies, claims, payments, and interactions.
Usage, billing, and support events reveal where every account is heading, long before renewal.
Building an internal model means hiring 3 to 5 ML specialists at $200K to $350K each, finding the right team, and often paying another $100K+ in recruiting fees. With Lyon, there are no new hires and no new department to build.
Before a model creates value, teams need data pipelines, training infrastructure, deployment systems, monitoring, governance, and security reviews. Lyon removes the need to build and operate that stack from scratch.
Maintenance typically adds another 25 to 35% of the original build cost each year. Lyon keeps models current as your data and business change.
Building an internal model takes months of scoping, implementation, testing, and deployment before teams have anything live. Lyon uses state of the art foundation model technology to turn proprietary data into production predictions in days.
Your data contains millions of possible relationships across customers, transactions, products, and time. No human team can manually discover more than a small fraction of them, especially the multi-step and temporal patterns that only emerge when the data is analyzed as a connected system.
When an internal model does not work out, you are left with a department to unwind, difficult staffing decisions, and no product in production. With Lyon, there is no permanent organizational bet. If it is not the right fit, you simply end the contract.
Every month without a working prediction model has a measurable cost. In emerging-market consumer lending, default rates of 5–8% are the norm. On a $10B book, that is up to $800M sitting in default. At a fixed approval rate, each point of KS lift screens out roughly one more percentage point of the bad accounts, and replacing a legacy scorecard typically lifts KS by 3–6 points, putting $20M–45M a year back on your books. Similar opportunity costs exist across fraud, churn, LTV, demand forecasting, and other prediction tasks.
Training and inference run entirely inside your VPC, on your cloud of choice. Nothing crosses your perimeter, not during training, not in production.
The foundation model is trained on your data alone and belongs to you. It is a proprietary asset no competitor can buy, rent, or replicate.
Access controls, audit logs, and reproducible training runs from day one. Your security team reviews every path data can take, all of them internal.
One company, one model. Your data is never mixed with anyone else's, never used to improve another customer's model. We never hold a copy.
Within a few years, every serious company will run on a model of itself. We're already training them.
Lyon was founded by a team from Stanford and NVIDIA with backgrounds in machine-learning research and mathematics. We spend our time on the frontier of self-supervised learning over structured data, and on making it work inside real enterprises.
That means we care about the unglamorous parts too: messy schemas, petabyte-scale event streams, security reviews, and models that have to earn the trust of the teams who bet on them.
It starts with a sample of your event data. Weeks later, you're reading predictions from a model that has learned your business.
Combinator