AI governance

System boundaries, evaluation, operational controls, and accountable delivery for applied AI.

AI governance for software systems: controls that fit delivery · Guide · A delivery-focused guide to AI inventories, risk boundaries, evaluation, human oversight, and operational evidence.
Architecture Solution Blueprint platform · Project · Designed and led the zero-to-one delivery of an enterprise platform that turns 1,000+ architecture patterns into real-time service recommendations, solution blueprints, and automated architecture workflows.
ASB Assist · Project · Building an LLM and RAG architecture assistant that turns natural-language requirements into architecture recommendations and diagrams using retrieval, tool integration, evaluation, grounding controls, guardrails, and auditable workflows.
Staff Software Developer & Enterprise Architect — Royal Bank of Canada · Experience · Building enterprise platforms and standards across architecture automation, observability, cloud infrastructure, and responsible AI.
"AI red teaming: a practical playbook for enterprise teams" · Note · AI red teaming is the practice of adversarially testing AI systems before attackers do. It borrows the name from cybersecurity red teams, and the analogy is useful but incomplete: you…
"Third-party AI vendor risk management" · Note · Enterprise AI systems are built on third party components: foundation model APIs, embedding services, vector database vendors, agent frameworks, evaluation platforms, annotation…
Design APIs for agent consumers · Note · APIs were designed for human developers reading documentation. AI agents are becoming a significant consumer of the same interfaces, and they read APIs differently: they parse schemas…
"Audit logging for AI systems: what to record and why" · Note · AI systems make decisions and take actions that need to be explainable after the fact: why did the agent do that, what data did it use, who approved it, what exactly happened? Audit…
Data retention and deletion in AI systems · Note · AI systems are data copies all the way down: training sets, vector indexes, caches, logs, evaluation sets, and backups. Each copy has its own lifecycle, and deletion requests — from…
"Defending against prompt injection in production LLM systems" · Note · Prompt injection is the defining security problem of LLM applications. It is not a bug that gets patched; it is a structural property of systems that mix instructions and data in the…
"Evaluating AI agents: beyond task success rate" · Note · Task success rate is the metric everyone reports and almost no one should rely on alone. It tells you whether the agent completed the task — not whether it completed it well, safely,…
"Guardrail architecture for LLM applications" · Note · Guardrails are the controls that keep LLM applications within acceptable bounds: blocking disallowed content, enforcing policies, validating outputs, and constraining agent actions.…
Incident review culture for AI systems · Note · Incident review for traditional software asks what the system did wrong. Incident review for AI systems must also ask what the system did confidently, plausibly, and incorrectly — and…
Model risk management for LLM systems in regulated industries · Note · Regulated industries — banking, insurance, healthcare — have managed model risk for decades. Credit models, fraud models, and capital models all pass through validation functions that…
"PII detection and redaction in AI data pipelines" · Note · AI systems are PII magnets. They ingest documents full of personal data, process user inputs containing identifiers, generate outputs that recombine sensitive information, and log…
RAG failure modes in production and how to contain them · Note · Retrieval augmented generation fails in production in ways that never appear in a demo. The demo uses a clean corpus, a cooperative user, and a forgiving evaluator. Production…
Production RAG requires retrieval evidence and control · Note · Retrieval augmented generation is often described as “search, then prompt.” A production RAG system is more accurately a governed information retrieval system connected to a…
Design safe tool use for AI agents · Note · An AI agent becomes materially different from a chatbot when it can read private data or change external systems. The central safety question is not whether the model produces good…
Build AI governance into the delivery control plane · Note · AI governance fails when it exists only as a policy document reviewed before launch. Controls must be attached to the same delivery path that versions models, prompts, data, tools,…
Evaluate LLM systems as systems · Note · A language model feature cannot be evaluated by trying a few prompts and deciding the answers look good. Its quality depends on the model, instructions, tools, retrieval, application…
Define the boundaries of a production RAG system · Note · Retrieval augmented generation connects a language model to external information, but it does not automatically make an answer correct, current, or authorised. A production design must…