AI adoption · evaluation · strategy
I build AI systems, then prove they work.
Extraction pipelines, adoption decision tooling, production reporting, generated content with automated screening. Two were built inside paid client engagements — a reporting pipeline used from 2025 to March 2026, and a certification framework handed over when I left. The rest I built and evaluated myself, and each entry says which it is. Twelve years of quality leadership behind it, which is why the interesting part is usually the evidence, not the demo.
- 12+ years
- in quality and delivery leadership, from sole in-house QA to QA Manager
- 70+ testers
- coordinated through team leads in the largest organisation; 45–50 with direct hiring and performance authority
- 4 platform families
- release-quality scope across mobile, web, TV and console, including payments and certification
- Since 2025
- building and evaluating AI systems hands-on, two of them inside paid client engagements — with seven case studies that set out what failed on first contact and the numbers behind the fix
Selected work
Selected systems and practice
Vision-to-catalogue extraction pipeline
Photograph a physical item, get back the verified catalogue record.
One case left deliberately failing: the record does not exist in the source catalogue
- n8n
- Vision LLM
- Cloudflare D1
AI use-case decision-support platform
Decision-support software for choosing which AI use cases an organisation should fund and ship.
Around ten design decisions documented with the alternative rejected
- React
- TypeScript
- Zustand
Offline certification framework with LLM-generated content
Desktop exam software, plus 650 bilingual assessment items generated and validated by language models.
Handed over when I left, at my manager's request. I have no data on how many people sat it afterwards, and I am not going to estimate — what I can claim is the build and the generation system, not the adoption
- Electron
- Node.js
- Python
Daily LLM reporting pipeline, in production
Test metrics, stakeholder summaries and risk reports generated every morning from live delivery data.
Built inside a confidential client engagement — client and product are not named
- LLM API
- Delivery tooling integrations
A personal AI operating system
The infrastructure I run my own work on: memory, tickets, sync, budget rules, generated views.
Enforced cost discipline: exploratory fan-out requires stated scale and approval
- Node.js (zero-dependency)
- Git
- Claude Code skills
Image classification for a small retail business
A phone-first tool that sorts incoming second-hand stock into categories from photographs.
Awaiting the customer decision. Dedicated API billing, authentication and accuracy tuning are scoped but not built
- Next.js
- Vercel
- Vision model
The reasoning behind it
Five questions, in this order
Starting at question three is how AI spend gets wasted. Stage four is the one that most often has no owner — and it is where I come from.
- 1 Find the bottleneck Business diagnosis
- 2 Choose the instrument Judgement
- 3 Redesign the process Process thinking
- 4 Prove it works Engineering rigour — the release gate between pilot and production
- 5 Make it stick Enterprise scaling
Contact
Where I am useful
Deciding which AI use cases are worth funding. Setting the acceptance bar before a pilot starts. Getting something out of pilot purgatory — or killing it honestly.