Sovereign AI inference for research computing.
How HELIX CPU inference pods eliminate GPU dependency for AI workloads — validated at UNSW Sydney's Katana HPC and running live on IBM HPC Fusion at the University of Queensland.
UNSW's Katana HPC cluster operated with chronic GPU oversubscription while its Intel CPU capacity sat idle. Standard inference workloads were consuming GPU cycles needed for training and simulation. By migrating inference to HELIX CPU pods, UNSW achieved immediate capital avoidance and freed GPU capacity for its highest-value workloads — without purchasing additional hardware.
The return.
Deferred purchase of 25–40 A100/H100-class GPUs for inference workloads.
Power, cooling and licensing cost avoidance per year.
Deferred data-centre upgrades: racks, chilled water, power circuits.
Pilot success criteria met within a 6-week deployment window.
Estimated avoidance/savings from the UNSW pilot business case. M
What HELIX delivered.
CPU-native inference — no GPU required
HELIX deploys as a CPU pod alongside any existing application (OpenShift, Docker, Kubernetes, HPC Singularity), achieving 24–60 tok/s on commodity CPUs with 4 concurrent slots — architecture-agnostic across Intel and AMD. No data egress, no cloud API dependency, no GPU procurement.
Real-time hallucination prevention (CORTEX)
CORTEX scores every generated token against a pre-computed truthfulness manifold during inference — not after — and produces a verifiable, per-token audit trail in every API response. Hallucination is prevented at the token level, before it reaches the output.
Drop-in OpenAI-compatible API
A fully OpenAI-compatible chat-completions endpoint with tool calling, structured JSON and chat templates. Integrates with LangChain, LangGraph, AutoGen and any existing agentic framework — zero code changes.
Deterministic decisioning (EAV-DT)
For structured decisioning from unstructured data — student risk, research-grant prioritisation, compliance screening — the EAV-DT delivers deterministic, auditable outputs with a mathematically guaranteed 0% decision-flip rate. Validated at 88–99.8% accuracy vs a 29.9% GPT-4o baseline on the same task.
Benchmarked on AMD EPYC 9254 (256 GB DDR5); UNSW runs HELIX on Intel — HELIX is architecture-agnostic. M · see claim substantiation.
A scoped proof-of-concept on one workload.
HELIX deploys in under one day alongside your current stack. The UNSW pilot achieved its success criteria within six weeks and required no new hardware. Book a technical deep dive to scope a POC in your HPC or cloud environment.