Articles

Security bolted on at the end is theatre. Treated as a design constraint from the first boundary — threat-modelled early, safe path as the default — it's what actually reduces risk.

Buying everyone AI licences isn't a strategy. A real one decides where AI creates leverage, how you vet the tools, the guardrails around them, and how you'll know it worked.

The core technical-leadership skill isn't picking technologies — it's converting vague, shifting business intent into a direction clear enough to act on and cheap enough to change.

When a team grows, decision-making breaks before the code does. Scale decisions — encode judgement into standards, paved paths, and culture so autonomy stays safe.

LLM coding tools amplify the engineer using them — most useful when you already know what you want, and quietly corrosive where a team leans on them to think.

The demo is the easy part. Production LLM systems live or die on evals, guardrails, and the failure modes that never appear in a demo — engineer the system around the model, not the prompt.

The build-vs-buy binary misses the option most teams actually take — bend. A framework for deciding by constraint and reversibility instead of ideology.

A decision framework for session-based vs token-based auth — revocation, topology, and the cookie and CSRF details that decide it in practice.

Turning a manual PDF slog into an automated pipeline — with de-identification, k-anonymity, and data governance designed in from the start, not bolted on.

When running compute on Cloudflare while keeping identity and data in Azure is the right call — and the integration realities that decision brings.

Matching a photo to a fixed palette is a small lesson in a problem that shows up everywhere — approximating a continuous space with a discrete one, using a metric that models the human eye.