Methodology
AI accelerates the work. 24 years of hands-on craft is what lets me check its output. This page documents how I select, orchestrate, and verify AI tools across professional delivery.
Orchestration, not adoption
The differentiator is not "I use AI." Everyone uses AI. The differentiator is demonstrated orchestration and human-in-the-loop judgement: naming the tool, explaining the selection rationale, documenting where the model output was wrong, and showing how experience caught it before it shipped.
Every AI claim in my portfolio sits adjacent to evidence. Not "AI-native" as a bare assertion. The recurring proof beat is: tool chosen, why, what the model got wrong, how experience caught it.
Tool selection
I reach for different AI tools depending on the task, not a single tool for everything. The selection is deliberate:
Claude / Claude Code
Architecture reasoning, code generation, document analysis, content strategy. Claude Code for multi-file refactoring and codebase-aware edits. Where the reasoning quality justifies the cost and latency.
GitHub Copilot
Inline code completion, boilerplate generation, test scaffolding. Faster than Claude for line-by-line development within an established codebase. I use it where speed matters more than reasoning depth.
Figma AI
Layout generation, auto-layout suggestions, asset search. Useful for rapid prototyping within an established design system. Not a replacement for design thinking, but an accelerator within it.
Copilot 365
Tenant-resident data gathering, document analysis and presentation scaffolding within Microsoft 365. Used where the task requires access to organisational content the other tools cannot reach.
Where AI gets it wrong
AI tools produce confident, plausible output that is regularly wrong in ways that require domain experience to catch. Examples from my work:
Hallucinated metrics
AI will confidently generate performance metrics, cost savings percentages, and user satisfaction scores when asked to write case studies. Without verifiable source data, these numbers are fabricated. Experience catches it because you know which metrics you actually measured.
Overclaimed seniority
Given a career history, AI consistently inflates titles, scope, and ownership. "Contributed to" becomes "led." "Supported" becomes "directed." "Second of two leads" becomes "co-equal technical lead." A 24-year career with real progression catches this because the actual arc is more interesting than the inflated version.
Accessibility false positives
Automated accessibility tools (including AI-assisted ones) pass components that fail in practice. A contrast ratio passes the algorithm but fails for the actual user group. Keyboard navigation "works" but the focus order makes no sense. Manual testing against real WCAG 2.2 criteria catches what automation misses.
Code that compiles but misses context
AI-generated code frequently produces syntactically correct solutions that violate project conventions, ignore existing patterns, or introduce subtle regressions. Copilot will generate a component that works in isolation but breaks the design system's token pipeline. Experience with the codebase catches it; the compiler does not.
The proof is in the portfolio
This site itself demonstrates the methodology. It was built with Claude Code, GitHub Copilot, and Figma AI. Every generated output was reviewed against the grounding document, verified against source facts, and corrected where the model got it wrong. The design tokens were extracted from Figma via an MCP pipeline, not generated by AI. The case study content was written from primary source material, not summarised by a model.
Client delivery. The AI workflow project is the clearest example: translating a university client's design system into a reusable AI skill, choosing Claude Code for the agentic build, Copilot 365 for tenant data, and Claude for the runtime. The practice manager asked me to present the method to my peers. The Anglo American ML platform demonstrates designing interfaces between human expertise and machine capability, working alongside an Azure ML pipeline in a 70+ developer programme.
Personal R&D. The SRHS proof of concept applies the same orchestration stance outside client work: designing a three-layer IoT architecture that assigns different AI tools to different tasks based on their strengths.