Projects & Work
The adventures, and the work that mattered. Two companies started from nothing in a country I had never worked in, one funded and scaled across six markets, and one business model I am currently trying to change from the inside.
Two companies, nine years
Arrived in China speaking no Mandarin and spent nine years building two companies there. West Education became one of Nanjing's leading training schools. Sinoed won four government contracts across schools, government departments and businesses. The skills that mattered most were listening, adapting and earning trust. Those travel further than any market knowledge I picked up.
Co-founder · Non-Executive Director
Co-founded a machine learning adaptive learning company in 2013. Raised multi-million euro investment across several rounds. Scaled to six international markets: Ireland, Singapore, Belgium, Spain, Slovakia and Germany. Ran every function at one point or another. Nine years from idea to operating business. Still on the board as Non-Executive Director.
Leadership team · Two years in
Trinzo sells hours, and AI is reducing the hours the work requires. I identified that before it hurt us, trained my own team, then took the business case to my peers on the leadership team. The market turned soon after and Trinzo contracted sharply, for reasons that had nothing to do with the programme, and we kept investing anyway. Two years in, on something that will take another three to five. Written with the confidence attached to each claim, and with the least flattering reading of it stated in full.
The reusable parts of the Trinzo AI Operating System, written up so they can be argued with rather than taken on trust. Two further pieces are in preparation.
The AI Front Door
Most organisations arrive with a long list of AI ideas and treat every item as though it needs the same governance. This sets out how candidates are sorted before any implementation method is opened: the four possible outcomes, what decides between them, and why routing everything into the controlled method is a failure rather than a precaution.
How the courses get built, and what happens when they meet a real organisation. The method first, then a course designed to it, then a programme delivered under an external regulatory deadline.
Method
The output of a training course is not the course. It is what people can do when it is over. Backward design applied to professional training: settle the capability, what must not deteriorate, what you believe causes learning and what evidence would convince you, before any content gets written. Each step ends with the check that tells you whether it actually happened.
Design account
A four-module Masters level course teaching regulatory professionals to decide whether an AI output deserves reliance, and to show why. An account of the design decisions rather than the outcome, because the course is built and has not yet run. The hard part was designing for a regulation that is not law yet: the volatile layer has to be swappable and the durable layer has to hold.
Delivered engagement
A large regulated organisation needed to update and scale an existing regulatory training programme for approximately 10,000 Quality professionals, against a fixed external FDA commitment. Three months of planned work was delivered in around three weeks, at roughly a seventh of the external cost, with every regulated output owned by a qualified person. The saving was not in the first draft. It was in every review cycle after it.
Engagements in regulated organisations, described at the level the client agreement allows. Each one is a diagnosis before it is a programme.
Life sciences · Two-track programme
Designed and delivered a two-track AI and regulatory compliance training programme for a large life sciences company. One track for all staff, one going considerably further for specialist QA and regulatory teams. The measure of success was whether people worked differently afterwards, which is harder to test for than whether they had learned the material.
Life sciences R&D · Adoption
Diagnosed why adoption of purpose-built AI tools stayed low in a large life sciences R&D department while generic tool use ran at around 75%. Designed a three-phase discovery, training and embedded adoption programme. Tools offered as optional aids get treated as optional. Adoption has to be built into the process itself, from the start.
Regulated manufacturer · Eleven weeks
An eleven-week programme for a regulated manufacturer whose quality, regulatory and lab teams were already using a general-purpose AI assistant daily, with no shared understanding of where GxP obligations began. Began with discovery rather than a framework: structured interviews, process mapping done on paper and walls, and an earlier external audit whose findings had never been turned into owned actions. Produced a risk-ranked process map, a role-specific boundary guide, traceability standards for documenting AI-assisted work, and a session training internal auditors to inspect it.
Device R&D · Instrumented pilot
A device R&D group wanted to know whether a general-purpose AI assistant could draft design-control documentation. Rather than survey every possible use, the pilot took one completed project and drafted a test protocol and its verification report, with the existing approved documents as the reference standard. Measured edit effort, error types, and whether review time actually beat writing time. The output was a decision rather than a demonstration: scale, iterate or stop.
The AI and LLM Implementation Playbook applied end to end to a named process. These contain no client engagement and no measured result. They exist so the method can be argued with.
Illustrative · Implementation Playbook
An illustrative worked example applying the AI and LLM Implementation Playbook to post-market complaint intake and triage. The interesting design problem is that complaint handling contains a genuine regulatory determination inside it. The reportability decision stays unassisted and the AI is confined to assembling the evidence the specialist needs to make it.
Illustrative · Implementation Playbook
An illustrative worked example applying the AI and LLM Implementation Playbook to drafting sections of a regulatory submission from approved source documents. Submission content is unusual in that fluency is genuinely dangerous: a well-written unsupported claim is harder to catch than a badly written one, and it goes to a regulator.