About Verificate
We build software that makes AI models work better — faster, more accurately, on the CPU infrastructure enterprises already own.
Building the Inference Layer for Sovereign AI
Craig Atkinson is the founder of Verificate, building the HELIX inference engine from Sydney, Australia. His background spans AI systems engineering, enterprise software architecture, and applied machine learning — with a focus on making large language models practical for sovereign, air-gapped enterprise deployment.
The core insight behind HELIX came from a fundamental observation: LLMs executing on CPU were performing far below their potential — not because they lacked capacity, but because irrelevant parameter activations were injecting noise into every token. The solution wasn't to make the model smaller. It was to make the execution more precise.
Craig led the design and implementation of HELIX v1.2 from first principles — the UTS telemetry layer, the truthfulness manifold, and the MoE precision slicing system — through to production validation on IBM Fusion infrastructure and the UNSW pilot deployment.
He has personally filed 3 US Provisional Patent Applications covering the core technology, established Verificate's IBM partnership, and delivered a completed UNSW pilot now in full deployment.
The Problem We Solve
Large AI models are expensive to run. GPU infrastructure is scarce, costly, and creates vendor dependency. Cloud inference exposes sensitive data and generates unpredictable per-token bills.
At the same time, every enterprise already has CPU infrastructure — AMD EPYC servers, OpenShift clusters, HPC environments. The hardware exists. The problem is quality: running a capable model on CPU, without sacrificing performance, has been an unsolved problem.
HELIX solves it. Not by shrinking models or degrading quality — by making execution precise.
Why Now
Three forces are converging in 2026: sovereign AI mandates are accelerating, GPU supply remains constrained and capital-intensive, and CPU hardware is advancing rapidly — NVIDIA's Vera CPU launch signals the era of CPU-first AI inference.
HELIX is built for this moment. The technology is production-ready, deployed, and validated. The market is arriving now.
Our two customer scenarios: enterprises moving from cloud/API to on-premises sovereignty, and those wanting to run larger, better models on the CPU hardware they own.
Company Milestones
Partners
IBM reseller partnership providing access to IBM Fusion HPC infrastructure, Granite model family, and enterprise distribution channels. HELIX v1.2 was benchmarked on IBM Fusion AMD EPYC 9254 hardware. IBM Granite 4.0 Hybrid Small is the reference implementation model.
University of New South Wales — pilot deployment complete and now in full production launch. HELIX deployed via Singularity/SIF on UNSW HPC infrastructure. Sovereign CPU inference running in production for research workloads. UNSW is both a partner and a live reference customer.
Get in Touch
Craig takes direct briefings with enterprise buyers, investors, and strategic partners. No sales team — direct access to the founder and technical lead.