invariant systems

agents

AI agents are heading toward full autonomy — writing, testing, and shipping software without supervision. For that to work, we need to trust the code they produce. Today, more AI-written code means more bugs and technical debt. This creates a bottleneck on human developers to review and fix the code.

We make AI-built software correct so it can become autonomous. As we work towards correctness, we build agents that are more productive and produce higher quality work compared to general-purpose agents.

Our technical foundation is OCaml — its strong compile-time guarantees, culture of testing and correctness, and credible path toward formal verification make it the right base for building AI systems we can trust.

We're also building the developer tools that make this possible — testing, benchmarking, linting, and eventually formal verification tooling — all designed for AI-first usage. The agent is only as good as the feedback loops it can rely on.

models

Generic frontier models aren't optimized for typed languages, build systems, or verification workflows. We fine-tune and evaluate models specifically for OCaml and correctness-sensitive development.

This, in turn, gives us a path towards local or self-hosted models that perform as well or better than generic models.

runtime

AI workloads are becoming a dominant share of computing infrastructure, and that share is only growing. At this scale, every layer of unnecessary complexity compounds — in cost, in latency, in attack surface. Today's stack wasn't built for this: general-purpose operating systems, container runtimes, and virtualization layers add overhead that made sense for general computing but not for a world where most machines run AI.

We're building infrastructure purpose-built for AI: models compiled ahead-of-time and deployed as self-contained binaries or unikernels, running directly on, or as close as possible to hardware. The result is faster, cheaper, and significantly more secure infrastructure.