Data stays in Canada
Canadian regions end to end — storage, compute, inference and backups. Residency is a deployment constraint, not a contractual promise about someone else’s cloud.
Canadian-owned · data resident in Canada
AI you can put in front of a regulator.
ShieldAI builds AI products for organisations that have to prove every decision — and designs the cloud, network-security and GPU infrastructure those products run on. Five platforms in production across translation, governance, intake, enterprise AI access and citizen services.
Engineered against
PIPEDA · PHIPA · FIPPA
ISO 27001 · SOC 2 · NIST CSF
IEC 62443 for OT
Official Languages Act · AODA
What every ShieldAI system shares
Four commitments, wired in at the architecture layer
These are not features we add per project. They are constraints the platforms are built inside, which is why the same assurances hold across a translation system, a GRC register and a GPU cluster.
Tamper-evident by construction
Hash-chained audit trails mean an altered record breaks the chain. You can demonstrate integrity to an auditor rather than assert it.
Human in the loop
Models triage, rank and draft. People decide anything consequential, and every routing decision is logged with the reasoning that produced it.
We own the substrate
The same team designs the fabric, the segmentation and the control plane underneath. Nothing critical is delegated to an assumption.
The portfolio
Five platforms, one engineering standard
Each one solves a specific regulated workflow. They share the same residency, audit and oversight guarantees, and they interoperate when it helps.
ShieldAI Gateway
Enterprise AI accessA compliance layer between your staff and the frontier models. PII is detected and protected before a prompt ever leaves your boundary, requests stay on Canadian infrastructure, and every call lands in an audit trail.
Elpista
Governance, risk & complianceNine GRC disciplines in one system — third-party risk, supply chain, cyber-maturity, evidence, control testing, risk registers, policy lifecycle, reporting and multi-framework mapping. Unusual depth in OT and industrial control environments.
Visit elpista.com
Licence-free vendor portal
Terminova
Translation managementThe full bilingual lifecycle — AI translation, centralised terminology, translation memory, revision workflow, SLA tracking, vendor portals and financial reporting. Git integration extracts strings and opens the pull request for you.
Visit terminova.io
Azure Canada East
WorkCohorts
Document intake & reviewAn AI agent for teams buried in paperwork. It reviews applications against your requirements, then auto-files the clean ones, flags what’s missing and escalates the high-risk cases — every decision logged and explainable.
Visit workcohorts.com
Your data never trains a model
Ontario Energy Assistant
Citizen servicesHandles the repetitive share of programme enquiries — eligibility, application status, documentation — so wait times fall and human staff keep their attention for the complex and sensitive cases.
Not sure which fits?
Describe the workflow and the regulator you answer to. We’ll tell you which platform applies — or say plainly if none of them do.
Infrastructure & architecture practice
We build the layer most AI vendors rent
Sovereign cloud, network and security architecture, and the tuned GPU fabrics that make large-scale training and inference actually perform. Engaged standalone, or as the foundation under one of our platforms.
Sovereign cloud architecture
Landing zones designed for residency and auditability from the first subscription — not retrofitted once a privacy review fails.
Azure and AWS Canadian regions, multi-region resilience
Identity, key management and policy-as-code guardrails
Evidence generated continuously, not assembled at audit time
Network & security architecture
Segmentation that survives contact with an incident, and designs that hold up when IT and OT have to share a boundary.
Zero-trust segmentation, micro-perimeters, east-west policy
IT/OT convergence to IEC 62443 zones and conduits
Controls mapped to NIST 800-53 and ISO 27001
AI infrastructure & GPU fabric
Where most clusters lose their throughput: the network. We design and tune the fabric so collectives run at line rate instead of stalling.
Lossless RoCEv2 fabrics — PFC priority queues, headroom and watchdog
ECN / DCQCN marking thresholds tuned against real congestion
Rail-optimised leaf/spine topology, NCCL collective profiling
Capacity, thermal and power planning for sustained training loads
FabricCheck — independent assessment of a cluster already running
Two ways in. Size it before you buy, and verify it once it is running.
FabricCheck
An independent performance check for NVIDIA GPU clusters
You bought the GPUs. Are you getting what you paid for? FabricCheck is an assessment of a cluster that is already running, carried out by engineers who did not sell you the hardware and have no reason to tell you it is fine. The faults that cost the most throughput rarely announce themselves — nothing errors, nothing alerts, the jobs simply complete slower than the invoice implies.
Links not running at rate
Ports negotiated below their rating, flapping links, error counters quietly
climbing, MTU mismatches, and PFC or ECN settings
that disagree between the switch and the NIC.
GPUDirect RDMA that is not actually on
The most common silent fault. Traffic falls back through host memory and
everything still works, just slower. Usually a peer-memory module that never
loaded, PCIe ACS left enabled and blocking peer-to-peer, or a NIC
sitting on a different root complex from the GPU it serves.
Servers dragging the rest down
Collectives are synchronous, so the slowest GPU sets the pace for every other one. A single node with a downgraded PCIe link or poor thermal behaviour costs you across the entire cluster, and averages hide it.
What you get: a plain-English report your procurement team can read and your engineers can act on — what is wrong, what it is costing you in throughput, and a fix list in priority order.
Request a FabricCheckOr start with the sizing calculator
Free, and nothing to install. Work out what serving an open-weight model from Canadian infrastructure actually costs — GPU count, memory, latency, topology. Every figure shows its arithmetic, including the cases where the honest answer is that you need less than you thought.
Compliance posture
The frameworks our work is built against
Canadian privacy law, international security standards and accessibility obligations — treated as design inputs rather than a checklist applied at the end.
PIPEDA
PHIPA
FIPPA
ISO 27001
SOC 2
NIST CSF
NIST 800-53
IEC 62443
AODA / WCAG 2.1 AA
Official Languages Act
French Language Services Act
Framework alignment describes how our systems are designed and assessed. Certification status for a given platform is confirmed in writing during procurement.
How we work
Short path from problem to something running
Modular deployments configured in weeks rather than quarters — because the compliance groundwork is already done and doesn’t have to be rebuilt for every engagement.
1
Scope against the regulator
We start from the obligation you actually have to satisfy — the statute, the framework, the auditor’s question — and work backwards to the system.
2
Pilot on real work
A bounded deployment against live workload in your Canadian tenant. You see genuine throughput and genuine failure modes, not a sandbox demo.
3
Scale with the evidence
Expand once the audit trail, the accessibility review and the security architecture have all held. The evidence pack comes with it.
Bring us the constraint everyone said was blocking.
Residency, auditability, bilingual obligation, an OT boundary, a GPU cluster that won’t scale. Tell us what it is and we’ll tell you straight whether we can help.