Articles

Applied AI notes, case-based analysis, and practical reflections grounded in real operational problems.

The Sandbox Problem — What Happens When Your AI Doesn't Share Your Reality

Why LLMs operate in an isolated sandbox with no access to real-time data, how missing context produces plausible but wrong output, and why closing the gap is a system design problem — not a prompt engineering problem.

Stop Paying Senior Rates for Junior Work: Model Delegation in AI Coding

How to split architectural decisions and routine implementation across different AI models — and why your PRD.json makes it possible. Use your strongest model for decisions that shape the system and a faster model for everything else.

Stop Your AI Agent from Hallucinating: Use a PRD.json

How a simple JSON file can turn chaotic AI coding sessions into structured, step-by-step execution. Learn the PRD.json pattern to eliminate agent drift, prevent hallucinated code, and keep your AI coding tools focused on what actually matters.

Software as a Living Structure of Thought

Three ideas from Umberto Eco — teaching judgment over information, building living structures instead of passive archives, and using language to project possibility — converge into a single argument about the future of software and education in the age of agentic AI.

AI for Internal Bug Reporting: A Better Way to Create Useful Tickets Without Losing Human Control

How an AI triage agent can transform vague internal bug reports into structured, actionable tickets — without automating away accountability. A practical model for small teams that need process without platform overhead.

AI-Written Code Still Needs Serious Testing After the First Commit

Why post-commit verification is a critical engineering phase for AI-assisted development. A real-world debugging session reveals how stacked bugs in data pipelines only surface through structured, domain-informed investigation.

You Don't Always Need to Train a Model: Building a Self-Improving Image Recognition Pipeline with Claude Code

How a local-first Python pipeline reached 85% accuracy on real-world product recognition without training a custom model — using test-driven optimisation, Claude Code as the iteration engine, and disciplined evaluate-hypothesize-modify-retest loops.

Historical Data Is the Real Foundation of Useful AI

Why historical data, not models, determines whether AI systems produce real value in business environments. A practical analysis of data quality, operational context, and the gap between demos and production.