AI that runs inside the hospital, not across the border.

A sovereign medical AI layer: the AI runs inside the hospital, on hospital-controlled hardware, and stays auditable.

• Models, memory, and data pipelines run on the customer network. Sensitive patient data stays local.
• Every output is grounded in authorized clinical sources and is fully traceable for governance and liability review.
• Pricing is a predictable site license, not a cloud token meter.
• We do not ask hospitals to trust a cloud vendor's privacy pledge. We remove the need for that trust.

We are not a governance dashboard and we are not a cloud scribe. We are the layer that runs the AI and proves what it did.

Sovereignty by architecture, not by contract clause.

• Patient data is de-identified and protected locally; differential-privacy budgets are tracked and audited.
• Synthetic datasets enable multi-site research without exposing re-identifiable records.
• No reliance on cloud-vendor privacy pledges or cross-border legal exemptions.
• This is the structural answer to PHIPA, HIA, FIPPA, Law 25, and PIPEDA at once: minimize, localize, and prove it.

The Canadian privacy-preserving ML research base is real (CHEO/uOttawa synthetic data, Waterloo differential privacy) but rarely in routine clinical use. Oction productizes it.

Every clinical record becomes queryable on local hardware.

• Local vectorization across policies, protocols, formularies, and reports.
• No external API calls, rate limits, or service interruptions.
• Scales from a single clinic to a provincial health authority.

Honest status: target average query latency is sub-second on commodity on-premise hardware (target, not yet a measured clinical-corpus result). A baseline benchmark on a representative clinical corpus runs in Month 1 of the first health-system pilot.

Healthcare cannot deploy AI it cannot audit.

• Answers cite the authorized sources they are grounded in.
• A domain ontology constrains output to medically appropriate concepts.
• Every output writes an audit trail for clinical governance and liability review.
• A built-in eval loop monitors drift, accuracy, and safety over time.

Honest status: source-citation rate and hallucination reduction are target metrics, validated in the first pilot against a baseline LLM using a domain eval suite of 100 to 200 clinical question-answer pairs. We report the measured numbers, not the marketing numbers.

The same local layer powers the full clinical AI roadmap.

This is not a slideware company. The architecture is running in a regulated Canadian environment today.

• Phase 0 deployment is livethe Cyntech Group geotechnical knowledge assistant runs on Oction's on-premise stack in a regulated Canadian enterprise.
• The same architecture targeted for health systems is operational now: local LLM proxy, memory API, vector store, and cache, all on controlled infrastructure.
• A NIST 800-53 security baseline is implemented, documented, and audit-ready.
• Healthcare is the next vertical, not the first attempt. The on-premise pattern is already proven in production.

A measurable 90-day pilot, designed to prove sovereign AI on a health system's own terms.

What a partner gets:

• A sovereign AI layer running on their hardware, their data never leaving the premise.
• A measured benchmark reportlatency, source-citation rate, and hallucination reduction on their corpus.
• No cloud lock-in and no per-token meter.

How a pilot runs:

• One clinical use case (policy and protocol retrieval is the recommended start).
• A data-governance and clinical sponsor on the health-system side.
• A representative corpus to benchmark against.

The point of the pilot is simple: show that private, auditable AI works inside a Canadian hospital, measured on real clinical material.