AI at Full Depth: From Daily Practice to Operating Model

Generative AI has made producing work almost free. It has not made trusting that work any cheaper. A model will write a specification, a proposal or a thousand lines of code in minutes, and any of them can be confidently wrong.

Most organisations respond in one of two ways. Some bolt a chatbot onto an old process and get more output with the same bottlenecks in review, judgement and accountability. Others hold back and fall behind the teams that move. Both treat AI as a tool to adopt. It is a new way of building, and it needs its own operating model.

The question I set out to answer: how does one leader use AI across every part of the job, from research and strategy to code, documents, design and enterprise deals, at a speed traditional teams cannot reach, without lowering the standard of what ships?

I know AI inside out, from the model to the operating model, and I use it in every part of the job. Seven disciplines carry it.

1. Deep knowledge of how AI works, layer by layer.

• Models. Frontier and open models, what each is good at, where each fails, and when a small local model beats a large hosted one on cost, privacy or speed.

• Context engineering. What a model must see to get the answer right: the raw source rather than a summary, structured so the evidence survives the context window.

• Retrieval. Search, embeddings and grounding, so answers come from the business's own records rather than the model's memory.

• Tools and MCP. Safe, scoped access from models to real systems through tools and the Model Context Protocol.

• Agents and orchestration. Multi-step agents, sub-agents working in parallel, checkpoints between steps, and the exact point at which a person must decide.

• Evaluation. Test sets, error rates and failure modes measured before launch, because AI readiness is a rate claim, not a demo.

• Economics. Cost, latency and reliability priced per completed task, not per answer.

• Safety and governance. Guardrails, data boundaries, prompt injection risk and audit trails, so a team can trust what ships.

• Vision and multimodal. The image understanding behind catalogue enrichment, AI skin analysis and virtual try-on.

2. A multi-agent stack, orchestrated daily.

• Coding agents. Claude Code and Cursor work as my engineering team, often several agents in parallel on one machine, each on its own repository and task.

• Agentic workspaces. Agents connected to mail, documents, meeting notes and CRM data, so research and analysis start from the real record rather than from memory.

• Local models and infrastructure. Open models through Ollama, and containers and Kubernetes on my own machine, so agent-built services run and are tested locally before they ship.

• Rapid interfaces. AI app builders such as Lovable for front ends, so a working screen exists before the first design review.

3. Products I built myself with AI agents.

• In the last few years I have built my most recent products myself with AI agents: a payment reconciliation engine, a jewellery ERP, returns and exchange apps for Shopify and Adobe Commerce, and the portfolio platform this page runs on.

• They stand on eleven years of building products the classic way: more than 20 products designed and delivered end to end with engineering teams. That foundation is why I know what good looks like before an agent writes a line.

• Every repository carries its own agent constitution, written in CLAUDE.md and AGENTS.md files, that fixes the language, testing, style and review rules every agent must follow. The standard lives in the repository, not in my head.

4. Expertise turned into reusable agent skills.

• An enterprise document engine that produces board-ready documents and passes four automated quality gates before a person reads a word: structure, banned characters, repetition and layout.

• A presentation engine that learned a brand's visual language from its master deck and now builds client decks in that language.

• A design system packaged for agents. Tokens, components, page recipes and agent instructions, so every agent-built website stays on brand instead of reinventing the design.

• A competitive intelligence skill that monitors rivals across every product module and keeps the analysis current.

• An editorial skill that encodes my standard for long-form writing, so every essay meets the same bar.

5. AI for research and strategy.

• Competitive teardowns and gap analyses across every product line, mapped against the category leaders in each segment.

• Market maps of emerging channels such as quick commerce, built from primary sources and checked claim by claim.

• A manufacturer's 21-sheet production workbook turned into a complete manufacturing execution system specification, with lot tracking through every operation.

• Pipeline and revenue analysis across thousands of CRM records, computed in code and reconciled to the source, never estimated in prose.

6. AI inside the products I lead.

• Catalogue enrichment at production grade: taxonomies, attribute schemas, bilingual English and Arabic content, quality control pipelines and search-ready product data.

• The science of AI skin analysis and recommendation behind virtual try-on, documented as a full methodology paper for the beauty brands that use it.

• The product vision for an autonomous commerce intelligence layer: a composable weekly intelligence product that reads across an entire commerce suite from one unified data lake.

7. Verification designed before generation.

• The evidence an agent must produce is defined before the agent starts.

• Every figure is recomputed in code, every claim is traced to its source, and high-stakes work gets a second, independent review.

• A person owns the question, the check and the decision. Agents own the volume.

AI is not a side skill for me. I understand it from the model up, I build with it every day, and I design how teams use it.

Work that traditionally needs separate research, engineering, design and documentation teams now runs through one AI-native workflow that I designed, with the checks built in rather than added at the end.

My most recent products, a library of agent skills in daily use, and enterprise proposals, specifications and working prototypes all come out of that workflow, at a speed that changes what a single leader can take on.

Winner of the company's Generative AI Crackathon in 2023, and editor of The Rigorous Builder, a journal of long-form essays on AI, product and commerce systems.

The principle behind it: AI makes output cheap, so the scarce skill is judgement. I design systems where agents do the volume and people own the trust.