Evaluation harnesses, escalation policy, data lineage and the argument about who is accountable when a system acts alone. That is where deployments are won, so that is where we spend our time.
San Francisco / Amsterdam / Singapore
2019
// Founded
48
// Systems live
62
// Team
3
// Locations
In 2019, four of us kept arriving at the same post-mortem. A capable model, an enthusiastic sponsor, a pilot that impressed the room — and then nothing. The work died somewhere between the notebook and the production database.
01 / The diagnosis
The cause was almost never the model. It was permissions that could not be carried through to an answer, data that was clean in the sample and chaotic at the source, and an organisation with no agreed answer to who is accountable when a system acts on its own. Those are engineering and governance problems, and they are solvable.
02 / The response
So RYVEN was built around the unglamorous half. Evaluation harnesses, escalation policy and data lineage get as much of our time as model selection. Every system runs inside the client’s own boundary, because that is what lets legal and security say yes, and we agree the numbers it must move before a line is written.
03 / Seven years on
That discipline is now the whole business. Forty-eight systems run in production across healthcare, banking, energy, retail and logistics. Our first client is still with us, except their team now ships without our help, which we have always considered the correct outcome.
48
// Systems in production
3
// Research locations
62
// Engineers and researchers
7yr
// Average client tenure
We want AI to be as unremarkable inside a business as a database or a payment rail: dependable, owned, monitored and boring in the best sense. That means building systems people trust enough to stop thinking about.
The next decade of competitive advantage belongs to organisations that hold their own data, train on it privately and act on it in real time. We are building so that this is available to any operator with a real problem, not only to the handful who can staff a research lab.
We have never counted milestones in funding rounds. These are the moments where the practice actually changed.
RYVEN started as a research collective asking why almost every enterprise AI pilot died before production. The answer was rarely the model. It was the wiring around it.
A freight operator let us route live exceptions through an agentic system. It held through peak season, and the methodology behind it became our delivery playbook.
We shipped our own deployment runtime so clients could run fine-tuned models entirely inside their own boundary, closing the gap between what compliance allows and what teams need.
Evaluation, drift detection and red-teaming stopped being an add-on and became part of every engagement. If we cannot measure it, we do not deploy it.
RYVEN now runs production intelligence for healthcare, banking, energy, retail and logistics operators, with research teams in San Francisco, Amsterdam and Singapore.
Each of these cost us something to learn. They now decide which work we take and how we build it.
A demo proves a model can do something once. We only count work that survives real traffic, real edge cases and a real on-call rotation.
01 / 04Your data, your weights, your infrastructure. We design so that leaving us would be inconvenient rather than impossible.
02 / 04Every claim we make is attached to a number that was agreed in advance and instrumented on day one.
03 / 04Autonomy has a boundary. We are explicit about where it sits, and we build the escalation path before we build the agent.
04 / 04Most of this business is not model research. It is judgement about where to stop, what to write down and who gets woken up when something misbehaves at three in the morning.
A RYVEN squad is four to six people and there is no junior layer running the delivery. The person who designs the architecture is the person who writes it.
We commit into your environment from the first week. There is no handover moment where a black box arrives, because nothing was ever hidden.
Every architectural choice comes with a short document explaining the alternatives and why we rejected them. It reads oddly like homework, and it has saved every project we have run.
If your data is not ready, or the process has no volume behind it, we will say so on the first call. Selling a system that cannot succeed is a slow way to lose a reputation.
Across San Francisco, Amsterdam and Singapore. These four set the direction.
Enterprise architecture, deployment strategy
Co-founder, Chief Executive
Model training, evaluation methodology
Co-founder, Head of Research
Inference runtime, platform reliability
Director of Engineering
Decision interfaces, human-in-the-loop systems
Head of Applied Design
Retrieval systems, data lineage
Principal Architect
Evaluation, drift detection, red teaming
Head of Assurance
Bring the process that frustrates you most. We will tell you honestly whether intelligence is the right answer for it.
// Every enquiry answered within one business day
