Generative AI Company in Edmonton

Edmonton GenAI budgets step up sharply once a regulated reviewer enters the picture. A scoped Edmonton GenAI proof of concept (one use case, ten thousand documents, RAG pipeline, evaluation harness, hosted demo) is a five to eight week engagement. A production deployment adds PIPA compliance, AHS or ATB cybersecurity review, and managed inference for the first six months. Multi-use-case rollouts across ATB Financial, AHS regional facilities, or a Government of Alberta ministry sit well above that once you include identity integration, retrieval index curation, and an ongoing red-team retainer. Edmonton sits below Toronto and Vancouver because we operate from HQ here without the Vector or Cohere talent premium. Fixed fee per phase, not open T and M.

AMII (the Alberta Machine Intelligence Institute) is one of three Canadian CIFAR AI institutes, alongside Vector in Toronto and Mila in Montreal. When an Edmonton client's problem is genuinely research-grade (novel reinforcement learning, planning agents, rare-event learning, alignment work) we scope a paid collaboration with an AMII fellow or a PhD student through the institute's industry sponsorship program. Our engineers handle the applied engineering, MLOps, evaluation harness, and productionisation work that most academic groups will not. For standard RAG, fine-tuning, classical NLP, and document extraction projects no AMII partner is needed. We will say which bucket your project fits in on the discovery call rather than oversell research depth we do not need.

Yes. We default AHS GenAI to AWS ca-west-1 in Calgary (launched 2024, closest AWS region to Edmonton) or ca-central-1 in Montreal, with Bedrock for Anthropic Claude and Cohere Command R+ both available in-region with documented data flow restrictions. If a workload must avoid US-controlled hyperscalers under contractual or political constraint, we self-host Llama 3 or Mistral on g5 or p4d instances in the same Canadian regions, behind AHS-managed network controls. PHI never leaves the Canadian region, every inference is logged with a deterministic ID for HIA audit, and an OIPC-style PIA is part of every project. We have shipped this pattern for two AHS-adjacent clinical decision support pilots.

The Artificial Intelligence and Data Act under federal Bill C-27 classifies AI systems as high-impact (credit decisions, employment screening, content moderation at scale, healthcare diagnosis, biometric identification, law enforcement) and imposes obligations around risk assessment, bias mitigation, transparency, monitoring, and incident reporting. For Edmonton clients: ATB and Servus credit decisioning copilots, AHS clinical decision support, and Government of Alberta benefits-eligibility models all classify high-impact and need a model card, bias audit, monitoring plan, and human-in-the-loop control. Internal productivity copilots usually classify standard. We run the classification during discovery and align documentation to whatever the final AIDA regulations require, which OSFI, FCAC, and AHS Cybersecurity already telegraph in current expectations.

Yes. Both ATB and Servus Credit Union face OSFI E-23 model risk management and FCAC fair-treatment obligations, which means a copilot needs version-controlled prompts, evaluation harnesses run before every production push, prompt-injection defences, deterministic PII redaction in both prompts and outputs, structured logging into their existing SIEM (Sentinel or Splunk), and challenger model documentation that internal audit recognises. We default to Canadian-residency models (Cohere Command R+ on Canadian endpoints, Anthropic Claude on AWS Bedrock ca-central-1, or self-hosted Llama 3 on Alberta GPU capacity). Output is gated by a guardrails layer running NeMo Guardrails or Llama Guard that blocks regulated-advice categories where a human agent needs to take the conversation.

Hallucinations: every retrieval-augmented system runs a citation-fidelity check where outputs must cite retrieved passages, and a separate evaluation model scores the citation against the source. We refuse to ship a system that cannot show its work. We tune retrieval recall before generation quality, because most apparent hallucinations are retrieval failures, not generation failures. Prompt injection: we sanitise tool-call inputs, never pass untrusted document content as system-role instructions, use NeMo Guardrails or Llama Guard for output filtering, and run an adversarial test suite (jailbreaks, indirect injection through retrieved documents, role-confusion attacks) before every production push. We assume any user-facing surface will be attacked, because they all are, and we have the test logs to show internal audit that we tested before launch.

Yes. Bilingual English-French is straightforward: Cohere Command R+, Anthropic Claude, and OpenAI GPT-4 handle Canadian French at production quality, and we run a separate French evaluation harness with Edmonton francophone reviewers (the Edmonton francophone community is the largest in Western Canada outside Quebec). For Indigenous languages (Cree, Dene, Stoney Nakoda) the situation is harder: foundation model coverage is weak, training data is scarce and culturally sensitive, and any deployment needs community consultation through the Indigenous Relations Office at the Government of Alberta. We have done discovery work on Cree-language scoping and we will be honest about what is feasible versus what needs partnership with First Nations technology groups before we propose a build.

Internal productivity copilot for ATB, AHS, or a Government of Alberta ministry: ten to sixteen weeks from kickoff. Week 1 to 3 is AIDA classification, PIPA audit, identity integration design, and retrieval index scoping. Week 4 to 10 is build, evaluation harness, and iterative tuning. Week 11 to 13 is red-team week and security review. Week 14 onward is staged rollout with shadow mode for the first cohort. Customer-facing chat at higher AIDA risk runs eighteen to twenty-six weeks because the regulatory documentation and human-in-the-loop design eat additional time. Self-hosted Llama or Mistral builds add three to five weeks for GPU procurement, inference optimisation, and load testing on ca-west-1 capacity.