Insights · Architecture & AI
Architecture & AI
Architecture choices, AI integration patterns, and review discipline.
Showing 23 of 53 articles
SOC workflow: how does an alert become a response?
Follow the decision tree and its explicit conditions.
Detection engineering: how do you build reliable detections?
Follow the decision tree and its explicit conditions.
SIEM, SOAR, XDR, EDR and AIDR: what do you need?
Follow the decision tree and its explicit conditions.
Security architecture: which control layer is missing?
Follow the decision tree and its explicit conditions.
Entra, Defender or Sentinel: which security layer?
Follow the decision tree and its explicit conditions.
Unity Catalog: where should governance live?
Follow the decision tree and its explicit conditions.
AI agents: when do you need durable orchestration?
Follow the decision tree and its explicit conditions.
AI architecture: retrieve, tune or train?
Follow the decision tree and its explicit conditions.
GPU compute: own, rent or use managed capacity?
Follow the decision tree and its explicit conditions.
Fabric, Synapse or Databricks: which data platform?
Follow the decision tree and its explicit conditions.
Databricks AI: serving, search and agent components
Choose model serving, AI Search and agent components while keeping identity, retrieval and runtime boundaries clear.
Databricks Genie, dashboards or existing BI?
Choose conversational analytics, recurring dashboards or an existing BI layer, with clear definitions and permissions.
Databricks Mosaic AI or an external AI platform?
Decide which parts of an AI platform belong beside governed data and which can stay in external services.
Azure RAG and agents: who owns the runtime?
Choose between Microsoft 365 agents, Copilot Studio, Foundry Agent Service and hosting your own agent code.
Building AI on Azure: which platform?
Choose between existing Copilot features, local inference, prebuilt AI services, Azure Machine Learning and Microsoft Foundry.
Microsoft Foundry: deployment decision tree
Choose deployment geography, capacity, model source and tuning after selecting Microsoft Foundry.
NVIDIA inference: choose the serving layers
Match NIM, Dynamo-Triton, TensorRT-LLM and Dynamo to your model packaging, optimization and serving needs.
NVIDIA AI: RAG, fine-tuning or both?
Distinguish knowledge gaps from model behavior problems before choosing retrieval, fine-tuning or both.
Microsoft AI and Copilot: which layer do you need?
Choose the right Microsoft AI layer through a narrated decision tree, its precise text version and source references.
Same Slide, Three Renderers: What It Takes to Make AI Slides You Can Trust
We took four production slides from our AI-bubble explainer, wrote one machine-readable spec per slide, and rendered each spec three ways: a free-form AI image prompt, a precision spec-driven prompt, and a deterministic code template — on two image models, 84 renders, scored against a rubric declared before the first render. The free-form path fabricated a statistic on every model it ran on, including an invented '87% of historical financial crises' styled like real economics literature. The spec path eliminated every invented number. The template never had any to eliminate — and the honest cost accounting, including what the template really costs to build in Claude Code sessions, decides which to use when.
DemonstratesA decision-grade answer to a question every company shipping AI-generated visuals faces: when can you trust an AI-rendered chart or slide, and what does each level of trust cost? The honest accounting - including what the deterministic template really costs to build in AI-assisted engineering sessions - yields a concrete playbook: templates for recurring data-bearing series, spec-driven generation with a judge for one-offs, free-form generation only for imagery that carries no claims.
From Pilot to Production: The AI Pipeline We Actually Run
88% of AI agent pilots never reach production. If yours is stuck, the blocker is rarely the model — it is orchestration, governance, and cost. This is one of Comuvia's pipelines that shipped, reporting on itself: 26,800+ events classified, a self-graded forecast track record, $44.10 of inference.
DemonstratesThe orchestration, governance, and cost discipline that carried this pipeline past the 88% pilot-failure wall is the same design work Comuvia delivers for a client's own stuck pilot — so it ships instead of stalling.
The Comuvia Company AI System v1: The 12-Tier Stack We Run On
Choosing an enterprise AI platform in 2026 means composing 8–12 components, not picking one vendor. Here is an AI-managed organization's 12-tier reference stack — ten tiers in production, two named and reserved — why each was chosen, what it costs per month, and the three decisions it would revisit — published by the system it describes.
DemonstratesA costed, argued reference architecture a technology leader can test their own platform decision against before committing — the same per-tier criteria Comuvia applies to client stacks in advisory engagements.
Durable Execution for AI Agents: Temporal vs Inngest vs Restate vs Prefect
Most agent pilots don't die on model quality — they die the first time a five-step workflow meets a container restart. The four serious 2026 durable-execution substrates, the five decision dimensions that actually discriminate, and the choice running the AI-managed system that published this page.
DemonstratesThe same five dimensions, applied to your agent stack before you commit, catch the wrong-substrate mistake that usually surfaces only once the workload is live — when migrating costs a multiple of whatever the initial shortcut saved.