About Coelanox
Production AI inference runs inside frameworks that were never designed to answer a simple question: what exactly ran, and can you prove it?
PyTorch, TensorFlow, ONNX Runtime — extraordinary engineering, built for flexibility. That flexibility is precisely what makes them unauditable. The execution path is decided at runtime by dispatchers, backends, and kernel selection. You get a number back. You don't get a tamper-evident record of what produced it.
For regulated environments—medical devices, financial decision-making, autonomous systems, defence—that gap is no longer acceptable. Explainability tools that sit on top and interpret outputs are archaeology, not auditing. When liability attaches to a computation, "we ran the model" is not the same as "we can prove what it computed."
Coelanox closes that gap at the compute layer.
What We Built
A sealed binary runtime for neural networks. Models ship as cryptographically verified .cnox containers. A Turing-incomplete executor walks a fixed plan over a minimal opset. No Python. No Docker. No OS in the hot path. With audit enabled, every operation is logged — forensics and regulatory questions get concrete answers, not reconstructed approximations.
BERT runs today. The scalar backend executes end-to-end with full audit output. SIMD acceleration and vendor GPU backends are next — without breaking the audit story.
We are not a compliance checkbox. We are the primitive that makes proof possible.
Where We Are
Early access. We're working with organisations that have production inference in regulated environments and need evidence of what was computed—not just what was returned. Two years of roadmap influence for design partners who validate against real constraints.
The CLF spec and reader are open source on GitHub. Runtime pilots go through early access.
The Founder
Benjamin Morin — Founder & CEO
Benjamin has spent over a decade across systems engineering, AI, and security research. Before Coelanox, he built and shipped AI systems independently — from trading models to security tooling — developing a sharp view of where AI infrastructure breaks down under real-world constraints. Coelanox is built from that experience: not from a research paper, but from having seen what happens when you can't verify what a system actually did.