Writing

How inventories create false confidence in AI decisions — a complete decision record is not the same as sufficient evidence. Companion to Evidence Debt.

The system was designed to operate, not to prove what it did.

One system, decomposed three ways: by function, by placement, and by authority.

An AWS whitepaper on modernizing election systems for security and scale.

Why APIs must be redesigned for autonomous AI agents as primary consumers.

Building a smoke-test system to catch what unit tests miss in AI-generated code.

Adversarial AI code review surfacing real findings that passed traditional testing.

Spec-driven development with AI agents versus vibe coding.

Building a knowledge platform with Claude Code — why hooks, not prompts, are the guardrails.

700+ controlled runs showing context, not just capability, determines agentic outcomes.

Observing how AI agents make decisions under controlled constraints.

Why effective human-AI partnerships depend on context, design, and treating AI as a teammate.

Applying Amazon's Working Backwards method to AI projects.

Comparing Claude and GPT models across response time, output, and user feedback.

Practical lessons from building with AI tools like Claude, Cursor, and v0.

Demystifying AI model types through restaurant analogies.

Accidentally building an LLM gateway while handling multiple foundation models.