writing

How I think about AI in software engineering.

Opinions, ideas, and working notes - from building complete multi-agent systems end to end and driving AI adoption across enterprise engineering organisations.

The enterprise second brain: distributed knowledge, centrally governed

Karpathy's LLM wiki became Google's OKF in eight weeks - the knowledge layer is being standardised. Why every large organisation needs a second brain: compiled by AI, validated by domain owners, governed centrally - and the accuracy, cost, and eval numbers that make the case.

read →

Building your own agent harness: AWS vs Google Cloud

Every production agent needs the same four pillars: identity, guardrails, cost control, and knowledge. How Bedrock AgentCore and Google's ADK + Agent Engine + Knowledge Catalog compare on each - and why you should rent the commodity loop but keep your contracts, evals, and audit spine portable.

read →

Automating regulated business processes with AI, end to end

One chargeback dispute, walked through an AI automation pipeline stage by stage - extraction, policy retrieval from a versioned knowledge layer, calibrated confidence, a decision-class routing table, and the audit record that answers "why did you file that?". Grounded in real tooling and real straight-through-processing numbers.

read →

The SDLC is becoming agentic - but the gate is the point

Coding assistants were act one. The real shift is agents drafting work at every stage of the lifecycle - and the organisations that win won't be the ones with the most autonomy, but the ones with the best-designed human approval gates.

read →

Measuring AI coding tools: feelings aren't findings

In one of the few rigorous studies we have, experienced developers believed AI made them 20% faster while it measurably made them 19% slower. What that perception gap means, and how to run pilots that produce evidence instead of enthusiasm.

read →

AI is an amplifier: rolling out AI to engineering teams

The research is converging on an uncomfortable truth: AI multiplies whatever engineering conditions already exist. Strong platforms and culture compound the gains; fragmented systems just produce chaos faster. A field guide to rollouts that stick.

read →

What building AI agents end to end taught me

I built a complete multi-agent platform - three cooperating agents, every layer from infrastructure to prompts. The lessons weren't about model choice: agents fail on context before they fail on intelligence, tools are API contracts, and guardrails are architecture - not an afterthought.

read →