Safety is the limiting reagent of industrial autonomy.
We are building the deterministic layer that lets AI agents drive factories without ever issuing an unsafe command.
SafeKernels is in a closed beta. Apply if this is you.
- You run a real cell with at least two robot vendors (UR, KUKA, FANUC, ABB, Dobot, Siemens, ROS 2…).
- You have an LLM-driven planner in production or a serious 2026 plan.
- A near-miss or rework event would cost six figures or trigger a safety audit.
- You have an engineering champion who can stand up a pilot in one cell.
- You're a single-vendor shop with no AI / LLM in the loop.
- You need a fully certified, SOC 2-stamped vendor today.
- You're a research lab without a real cell or hardware.
- You can't carve out four weeks for a guided pilot.
Building the deterministic safety layer for AI-driven physical systems.
SafeKernels began as a research question: how do you let an LLM drive a robot arm without ever producing an unsafe physical action? Every existing answer was either fragile (prompt constraints), limiting (tiny action vocabulary), or circular (the LLM self-certifying).
The prototype was built to prove a fourth option: a deterministic kernel that the LLM can never bypass. The safety case that came out of it turned out to generalise. Once you have a typed knowledge graph of the cell, multi-vendor coordination, audit trails, and cross-factory federation fall out for free.
That's the platform we're now productising.
Three principles. We live by them.
These aren't slogans. Each one is a tiebreaker we've used in a real engineering decision.
- 01Determinism over vibes.
If a safety guarantee can't be stated, it isn't one.
- 02Prove, don't promise.
Numbers carry units and test conditions. Claims come with code.
- 03The factory floor is the spec.
The engineer accountable for the cell has the final word on every release.
Be one of the first three deployments.
Apply
Tell us about your fleet, your hardest safety requirement, and the timeline you're working against.
Apply to the ProgramSafeKernels is pre-revenue at v0.1. We work hand-in-hand with the first cohort of factories deploying LLM-driven motion planning: founder-level access, joint architecture decisions, joint whitepaper and conference rights, and the ability to register custom safety patterns directly in the kernel.
What we ask
A real cell with at least two vendor robots, a written safety case the team has been struggling to close, and a willingness to publish results.
What you get
A 50% pilot discount on Starter for the first three months, weekly engineering review, and the technical reference architecture for your fleet.