Author:
I let the AI build a browser game on its own: 21 hours, 613M tokens and USD 240. It looks great. The code piled up 20 anti-patterns.
Keep reading: Blind vibe coding: a game that looks great and code that doesn't scale →
The media keep coming back to the same question: whether AI feels, whether it's conscious, whether it's afraid. Here I explain why an LLM is closer to a very complex weather forecast than to a mind — and why sounding convincing doesn't mean it thinks.
Keep reading: Is AI conscious? No: it calculates probabilities →
If all you do is publish articles, a full CMS is attack surface and maintenance you don't need. This is how I migrated a site to S3 + CloudFront using Claude Code, with minimal permissions, stage and prod pipelines, and no access keys in GitHub.
We use cognitive vocabulary to describe these systems, and that sets expectations that later break in production. Here I go over the basics — LLM, MCP, AGI — what really happens inside a transformer, why reasoning isn't a new module, and the Artificial Analysis metrics I use to measure instead of argue.
Keep reading: What an LLM really is — and how to measure it with Artificial Analysis →
AI dev tools are great for a PoC, but shipping to prod is another story. Here I share my approach to the AI SDLC and the best practices I apply, based on my own experience.
Keep reading: AI SDLC: a compendium of best practices for developing with AI →
A written companion to zerogap.uk/agents — where I share RAG architectures, agentic orchestration patterns and how to pick the right one for your problem.
One possible design for production-grade agent infrastructure on AWS: ECS Fargate, GitLab CI/CD with automated evaluation, Lambda ingestion, self-hosted Phoenix observability, a single RDS instance with pgvector, pilot-light multi-region DR and a continuous improvement loop — with monthly costs.
Keep reading: Agent Architecture — Production Infrastructure →
A pattern for letting SharePoint users work with back-end data without exposing secrets or trusting the client: an SPFx extension gets an Entra ID token silently, an Azure Function validates the JWT and writes to Azure Table Storage.
Keep reading: SPFx, Azure Functions and Table Storage with Entra ID →
The HyperAgent Cognition Engine is a dual hyper-agent architecture in which two complete cognitive pipelines — a Solver and a Critic — work together, combining lateral thinking, domain-expert reasoning, independent verification and iterative refinement in one coherent system.
Keep reading: HyperAgent — Cognition Engine: Building an AI System That Thinks Like a Human →
A practical checklist for securing AI systems: choosing between LLMs and deterministic logic (and when to use MCP), securing inputs, sanitising outputs, sandboxing agents and tools, using local models safely, and continuous evaluation and observability.
