Agents and the systems they run on — applied AI, distributed systems, and blockchain, engineered for production where being wrong is expensive.
Four levels of ownership — from a single decision to the whole system, run in production.
We run research in small, precise models — the part of the work that has to be understood before it can be relied on.
Task-specific models small enough to run on your own infrastructure, where the data does not leave your control.
Adapting models to a single domain and its vocabulary, so behaviour is narrow and predictable.
Systems that take actions, and the bounds and evaluations that make those actions safe to allow.
Building and curating the domain datasets that determine what a small model can do.
Verification methods for systems whose failures are quiet — where output looks plausible and is wrong.
A small team of senior engineers. We stay narrow because the failure modes in these systems are specific and hard-won — concurrency, consensus, models that drift — and knowing them is what separates a system that works from one that only looks like it does.