I'm a product manager in the UK civil service. My work sits where policy meets delivery: taking questions that arrive as documents and turning them into services that can be tested with the people who'll use them. Outside work I build software, mostly tools for the kind of problems I see in the day job but can't solve from inside it.
2026 – · Founder, product and engineering
A desktop app for macOS and Windows that turns a plain-English brief into accessible, design-system-conformant prototype pages — generated by a local model, with no code, cloud or account. Supports nine systems across six governments: GOV.UK (incl. MOJ overlay), NHS.UK, Scottish Government, USWDS, NSW, Bootstrap Italia, Japan's DADS, and a neutral Abstract system.
2026 · Builder (Frontline London Hackathon)
Captures post-trip debriefs from humanitarian convoy operators on the UK–Poland–Ukraine corridor and turns them into a living, dated knowledge base — the border intelligence that isn't written down anywhere. Every merged claim keeps who said it and when; the model extracts, the rules engine decides.
I keep large language models out of the decision path. They read messy input and write clear output; the judgement in between is deterministic, testable and versioned. AidRoute Debrief merges convoy reports through a rules engine with the model confined to extraction and narration. Prompt to Page's generation is wrapped in per-model quality gates and automatic repair of known failure patterns rather than trust.
Border intelligence, clinical evidence, generated HTML — if it can't be traced, it can't be relied on. Provenance is a design requirement, not a nice-to-have. In AidRoute every merged claim keeps who said it and when; in Prompt to Page every page ships with its own quality report.
Prompt to Page's NHS.UK output carried a "quality unknown" label until six models were measured at or above their GOV.UK level. For a chat product I built a three-layer hallucination evaluation — prevention audit, enforcement test, 43-case adversarial replay — and wired it into CI so the safeguard is pinned to the cases it must catch (10/10 caught, 0/8 false positives).
One validation call with a convoy organiser killed the original AidRoute concept and surfaced the real problem: unrecorded, word-of-mouth border knowledge. I run structured outreach to practitioners who use each design system Prompt to Page supports, and change the tool on what they report.
I work in the UK civil service, so my side projects are built on personal devices, declared, and non-commercial while I'm in post. Prompt to Page never bundles restricted fonts or protected government marks; CI fails the build if they appear. Knowing the boundary is part of the product.
Email hello@courtneyallen.co.uk. I reply to everything that isn't automated. More ways to reach me are on the contact page.