VW INTEL GROUP / LIBRARY

AI Infrastructure Library Canada: Runbooks, Case Studies, and Open Docs from Operator Engagements

Production work, written down. The AI infrastructure library Canada catalog publishes named research outputs from the VW Intel Group. Alongside those, deployment runbooks for Llama 3.1 on H100 hardware and case studies from Canadian regulated operators in healthcare and construction. Additionally, the catalog surfaces the public open-doc files that machines crawl directly. Furthermore, every methodology reproduces and every dataset versions on Hugging Face under a permissive licence. Moreover, every artifact downloads free rather than landing behind an email gate. Consequently, the library is the proof rather than the claim.

Subscribe to the Threat Brief →
Brutalist monochrome schematic of a single bound report viewed from above, with a red ribbon bookmark and an outline maple-leaf mark on the cover.

FEATURED ARTIFACT · Annual Report

Annual Sovereign AI Readiness Report 2026: How Canadian Operators Run AI Workloads Without Sending Data South

The inaugural annual benchmark of Canadian sovereign-AI readiness. Specifically, the report draws on the AI Infrastructure Exposure Scanner dataset and indexes every finding against PIPEDA articles and ITSG-33 control families. Additionally, it proposes four deployment patterns that allow regulated Canadian operators to run AI workloads without sending data south. Moreover, the methodology reproduces and the dataset versions on Hugging Face under CC-BY-SA 4.0. Finally, the regulatory mapping ships as a separate appendix for legal review.

Read the full report →

RECENT ENTRIES

Recent additions to the AI infrastructure library Canada catalog.

Schematic icon of a construction hard hat next to a clipboard, with a red chin strap and a single red document line.

Case Study

Case Study: BC Construction Firm Reduced Safety-Training Cost 73% with a Fine-tuned 7B Model

A Vancouver construction operator replaced a $1.8M annual safety-training programme with a Llama-7B fine-tune deployed on a single on-prem GPU. Specifically, we delivered a 73% cost reduction over 18 months. Moreover, the engagement added a Canadian-French capability tier for Quebec subsidiaries and preserved full audit-log compliance with WorkSafeBC training records.

Read full →
Schematic axonometric icon of a 1U server card module with a single red LED indicator and a cooling-fin pattern.

Runbook

Runbook: Deploying Llama 3.1 70B on a Single H100 with vLLM

Single-H100 deployment of Llama 3.1 70B served with vLLM under sustained production load on a single bare-metal Canadian operator GPU. Specifically, the runbook covers tokenizer settings, kv-cache memory sizing for 8k windows, and hot-swap recovery. Furthermore, it names the four operator decisions that determine one H100 fit.

Read full →
Schematic polar-projection routing diagram with one red dashed line crossing between two continent-edges.

Brief

Brief: Why We Route Canadian AI Workloads Through Hetzner Helsinki, Not us-east-1

Hetzner Helsinki keeps Canadian AI workloads inside an EU-adequacy jurisdiction while remaining PIPEDA-compatible for cross-border data flows. Specifically, this brief documents the network route, the bandwidth profile, and the full GDPR-PIPEDA interoperability rationale used by Canadian operators. Additionally, it cites the OPC’s published opinion on EU-routed processing.

Read full →
Schematic bar-chart icon with one tall black outlined bar and one short red filled bar, representing cost reduction.

Case Study

Case Study: Replacing $5K/month of OpenAI Calls with a 13B Fine-tune in 3 Weeks

A mid-sized Toronto SaaS operator was burning $5,000 per month on OpenAI completion API calls. Therefore, we fine-tuned a Llama 13B model in 3 weeks and consequently recouped the engagement cost inside the same quarter. Furthermore, the case study covers the per-token economics math and the three failure modes we hit during evaluation.

Read full →
Schematic axonometric icon of a stacked index graph with a single red node at the centre indicating retrieval.

Runbook

Runbook: Postgres + pgvector for a 2M-document RAG with Sub-100ms Retrieval

How we built a 2M-document retrieval-augmented-generation system on Postgres with the pgvector extension, reaching sub-100ms p95 retrieval at production scale. Specifically, the runbook covers HNSW index tuning and the embedding-store schema. Additionally, it documents the hybrid-search blend ratio between dense and BM25 lexical retrieval we ship in production.

Read full →
Schematic 2x2 grid icon of four deployment-pattern symbols with a single red highlight on the barrier-wall cell.

Brief

Brief: The 4 Sovereign-AI Deployment Patterns for Canadian Operators

Four named deployment patterns for Canadian operators, indexed by jurisdiction and data-residency requirement. Specifically, the Sovereign Box runs fully on-prem and the EU-Helsinki Pass-Through routes through Hetzner. Moreover, Sovereign-MCP Tooling federates models across mixed jurisdictions, while finally the Federal Protected-B Air Gap satisfies ITSG-33 control families for Treasury Board workloads.

Read full →

OPEN DOCS / MACHINE-READABLE

Files machines crawl directly.

Furthermore, the AI infrastructure library Canada catalog publishes its machine-readable spine at the root domain. Specifically, the llms.txt convention surfaces a Canadian-context content map for LLM crawlers. Additionally, the robots.txt advises polite crawl-rate guidance and the XML sitemap index enumerates every named output by canonical URL. Moreover, this open-doc spine is the surface that the Canadian Centre for Cyber Security and the Office of the Privacy Commissioner inspect when assessing third-party AI vendors.

External references · Canadian Centre for Cyber Security · Office of the Privacy Commissioner of Canada

READING GROUP

Reading more, deploying more.

The monthly Vanwebdev Reading Group convenes Canadian operators around a single paper, runbook, or threat brief from the AI infrastructure library Canada catalog. Moreover, rotating external researchers from Ottawa, Toronto, and Vancouver present alongside the host. Specifically, sessions are free to attend, recorded under a CC-BY licence, and transcribed for the archive within seventy-two hours.

Join Reading Group →