Private inference
Run representative models in controlled environments and measure performance, capacity and operating requirements.
Proposed Toronto reference capability
AFA Sovereign AI & Physical AI Reference Lab
A practical environment for developing and demonstrating private AI, model evaluation, institutional knowledge systems, accelerated computing, governed agents, simulation and Physical AI.
From concept to demonstrated capability
A practical Toronto reference environment
Atkinson Film-Arts proposes establishing a Toronto-based Sovereign AI & Physical AI Reference Lab built around Dell Technologies and NVIDIA accelerated-computing infrastructure.
The objective is substantially larger than acquiring a high-performance computer. The lab is intended to build practical expertise designing, demonstrating, deploying and supporting AI systems for organizations that need greater control over sensitive information, intellectual property, models, institutional knowledge or physical operations.
The hardware is the foundation. The value is the expertise, evidence and repeatable deployment capability built around it.
Reference Lab workstreams
Test assumptions instead of accepting them
The lab is intended to create a controlled environment for practical evaluation across the intelligence stack.
Model evaluation
Compare model quality, language, institutional assumptions, behaviour and version changes against the same workload.
Institutional knowledge
Test retrieval-augmented generation, authoritative sources, permissions, provenance and knowledge boundaries.
Model adaptation
Evaluate prompting, retrieval, fine-tuning and other adaptation methods where evidence supports them.
Agents & governance
Test tools, permissions, policies, logging, red-team scenarios, approvals and operational boundaries.
Physical AI
Connect accelerated computing to simulation, robotics, perception and real-world systems under bounded authority.
Canadian reference capability
Evaluate institutional alignment in practice
Canadian organizations should not have to assume that a globally trained model automatically understands Canadian law, institutions, terminology, geography, bilingual requirements or organizational context.
The Reference Lab can compare how models respond to Canadian institutional tasks, whether authoritative sources improve grounding, whether model adaptation changes behaviour and whether the resulting system can be documented, reproduced and governed.
Which AI system behaves appropriately for our organization, our jurisdiction, our responsibilities and the people we serve?
Representative lab questions
What organizations can test
The exact work depends on the customer and the maturity of the lab capability.
Where should this workload run?
Compare approved cloud, private, local, Canadian-hosted and hybrid placement according to sensitivity and performance.
Which model is appropriate?
Evaluate quality, domain performance, context, cost, latency and behavioural suitability.
Can we protect institutional knowledge?
Test permissions, retrieval boundaries, retention and controlled access to sensitive repositories.
Can policy become technical control?
Explore identity, data classification, model approval, agent permissions, logs and escalation.
Can we trust updates?
Compare versions and detect meaningful behavioural or performance changes before production adoption.
Can intelligence act safely?
Evaluate robotics, simulation and Physical AI with explicit operational authority and human oversight.
Customer journey
Assess → Architect → Demonstrate → Deploy → Govern
The Reference Lab supports a broader AFA services lifecycle rather than infrastructure resale alone.
Assess
Define workloads, information sensitivity, users, performance and institutional requirements.
Architect
Design model, knowledge, compute, networking, security and governance patterns.
Demonstrate
Use representative workloads to test assumptions and show stakeholders how the environment behaves.
Deploy
Implement an approved architecture at the appropriate customer scale and location.
Integrate & align
Connect institutional knowledge, applications, agents and sector-specific context.
Evaluate & govern
Monitor quality, behaviour, security, model changes, access and operational authority over time.
Continue the journey
Related pathways
Practical perspective
From understanding to practical action
Atkinson connects people, policy, infrastructure and implementation so each initiative can move forward with clarity, accountability and purpose.
Explore a Reference Lab Use Case
Bring a workload, model, research problem or Physical AI use case that would benefit from controlled testing.
Request a Lab BriefingIn practice
A governed system for coordinated intelligence
Each engagement is shaped around the organization, information, responsibilities and outcomes involved.