I’ve spent over 25 years in IT. Long enough to have stared at IBM 5250 green screens, watched data centres shrink from football pitches to single racks, and seen more technology revolutions than I can count.
This blog is where I think out loud about what actually matters in that change — and what doesn’t.
These days my focus is data and AI. Not the hype cycle version — the real one. How data strategy works in practice, what machine learning can and can’t do for a business, and why the foundations matter more than the headlines. I’ve spent the last several years studying Data Science and AI at Oxford, Imperial College, and I’m now working my MSc in Computing and AI — so I try to bring both hands-on experience and academic rigour to what I write.
Occasionally I’ll wander back into infrastructure and IT management. And yes, sometimes I’ll be found in the garage with my 1974 VW camper van and my 1979 MGB GT and no Wi-Fi in sight. Sometimes going analogue is the right call.
Views here are entirely my own. This is a personal blog — not a vendor platform. Nothing published here is approved or reviewed by any employer or technology vendor.
The presentation on agentic AI emphasized that successful AI implementation relies on foundational processes rather than solely on advanced models. Crucial elements include process documentation, data quality, defined permissions, and robust governance. Organizations must ensure readiness and controlled autonomy before deploying agentic AI, which requires careful planning and engineering, not mere assumption.
The article compares two flower classification models: a simple decision tree using the Iris dataset and a convolutional neural network (CNN) trained on images. The first model makes decisions based on measurable attributes, achieving high accuracy with minimal data. In contrast, the CNN analyzes pixel data, needing vast training resources. Both models have unique strengths, highlighting that the best choice depends on the problem at hand.
The author shares a weekend project of building a CNN image recognition system for identifying five flower types. The narrative transitions to Castellan Wines, where CEO Marco seeks AI for defect detection on a bottling line. Data scientist Clare emphasizes tailored solutions: classical methods for measurable tasks and deep learning for variable damage detection, highlighting engineering decisions over technology hype. Her blueprint outlines three detection strategies, underscoring the need for controlled environments and ongoing system validation.
The AI industry struggles with data quality due to systemic issues that prioritize modeling over data work. A paper highlighted that 92% of AI practitioners faced data cascades, resulting from cultural and educational frameworks that undervalue data integrity. To address this, the industry must reward data excellence and treat data as integral to AI development.
Do you remember your first magic show? Not on screen, but a live show where you’re in the audience just a few feet from the…
The author reflects on a principle learned early in their career: don’t over-engineer solutions based on assumptions of bigger is better. Data should dictate model size, as larger models don’t necessarily yield better outcomes. The importance of understanding data requirements is emphasized, particularly in the context of LLMs and operational costs.
