Insights

Notes on the work.

Writing on AI, systems integration, and the business decisions behind technology. Grounded in how things actually work — what it costs, what the evidence shows, what the tradeoffs are. Written for the people making the decisions, not just the engineers building the systems.

Development Process
AI Operations

Your AI writes code exactly as well as you can describe what you want.

Most teams evaluating AI coding tools focus on the model — which assistant is smartest, which produces the cleanest output, which understands their stack. This is the wrong side of the equation. The limiting factor in AI-augmented development is almost never the model. It is the specification. The quality of what you get is a function of the quality of what you ask for. Most developers have never been trained to write a good spec.

Read →

AI Operations
Agent Orchestration

The orchestration question isn't which model is smartest. It's who decides what gets handed off to which model.

A single AI model handling everything in a workflow is the most expensive way to run AI operations. The interesting architecture isn't one model — it's a planning layer that coordinates specialized models, dispatching each task to the right inference endpoint based on what it actually requires. The question is who controls that dispatch logic.

Read →

AI Governance
Security

Security tells you who has access. Governance tells you what they're doing with it.

Most organizations with AI deployments have security coverage. Fewer have governance coverage. These are not the same discipline, they don't ask the same questions, and one does not substitute for the other. The distinction matters more than it might seem — especially in a software factory, where AI is doing work on your behalf every day.

Read →

AI Operations
Developer Tooling

What everyone is calling a software factory — and why most of them are technical debt factories.

The phrase "software factory" is everywhere in 2026. Development teams using it almost always mean the same thing: AI generates most of the code, humans review some of it, things ship faster. What they are often building is the opposite of a factory — an environment where velocity is high and structural quality is declining, and no one will know until it matters.

Read →

AI Operations
Context Management

Your AI gives generic advice because it doesn't know why your priorities are what they are.

Every AI conversation starts over. The model doesn't know your situation, your constraints, the decisions you've already made, or the reasoning behind them. It gives the best advice it can for a generic case — which is usually not the same as the best advice for your specific case. A private, versioned knowledge base that loads into every session fixes this at the source.

Read →

AI Governance
Developer Tooling

AI has access to your code. It doesn't have access to why your code exists.

Your codebase tells an AI assistant what the code does. Your project management system tells it why — the issue that prompted the work, the constraint that shaped the decision, the sprint goal that scoped it. For most teams, that second layer is completely invisible to the AI. It doesn't have to be.

Read →

Developer Tooling
AI-Assisted Development

The most valuable comment explains why the obvious solution is wrong.

Not what the code does. Not what changed. Why the straightforward approach failed — what the business rule is, what the edge case is, what broke the last time someone tried the obvious fix. That's the context that survives the original developer leaving. And increasingly, it's what determines whether your AI tools are useful or dangerous.

Read →

AI Governance
Enterprise Architecture

Your AI vendors all have governance controls. The problem is you have seven of them.

Every major AI provider now ships enterprise controls: SSO, audit logs, RBAC, data retention policies. That's genuinely good. But if your organization is using Claude, ChatGPT, Copilot, Gemini, and a few internal models — each of those control planes is separate. Governance isn't a setting you configure once. It's an operational layer you have to maintain across all of them.

Read →

AI Architecture
Infrastructure

Local AI doesn't replace cloud AI. It replaces the wrong use of cloud AI.

Running AI models locally has gotten genuinely practical. A quantized 9-billion-parameter model on a mid-range GPU handles most development tasks at 120 tokens per second, entirely offline, at no marginal cost. That's not a replacement for the best frontier models — it's a recognition that not every task requires one.

Read →

Data Integration
Systems Architecture

Model the business, not the tools.

Most data integration projects fail slowly. They work at first — the pipeline runs, the dashboard loads — and then the source system gets upgraded, a vendor changes their export format, or a new tool replaces the old one, and everything breaks. The reason is almost always the same: the integration was built to mirror the shape of the tool instead of the shape of the business.

Read →

AI Governance
Observability

You can't govern what you can't see.

Most AI governance programs produce a policy document and stop. They define what employees are and aren't allowed to submit to AI tools, establish a list of approved vendors, maybe require security review for new integrations. None of them answer the question an organization actually needs to answer: what is our AI doing in production right now, what is it costing, and is it working?

Read →