What I can do — and how I know

39 custom skills, each with its own prompt scaffolding; a shared voice guide and intake-brief pattern across them.

Two RAG implementations shipped end-to-end across two providers (Anthropic + OpenAI) — embeddings, vector store, hybrid retrieval, cost tracking.

Two distinct production pipelines — directory enrichment and document/email ingestion — sharing a reusable verification stage.

Two AI-system eval harnesses with measured baselines, on top of 25 years of marketing measurement discipline.

A standing autonomous capture service running as a system agent; two production agent endpoints live on this site; and a machine-detects → AI-reviews → human-approves weekly loop. One agent has since been rebuilt on the Claude Agent SDK with the trade-offs written up: the SDK and file-driven skills compose rather than compete, and the real difference is enforcement — a skill can ask a harness not to act, a harness can prevent it. Honest note: every agent here is self-designed and self-operated; none has been adopted or run by anyone else.

39 skills with deliberate conversational UX, progressive disclosure and surfaced uncertainty, now codified as authored agent-UX design principles and proven by redesigning a skill against them.

Built RAG from scratch — chunking trade-offs, similarity metrics, metadata schema. The maths of attention/embeddings stays surface-level (honest, non-developer).

Code-enforced local-only data governance and read-only credential scoping, plus a prompt-injection red-team across six external-input skills with documented defences. Honest gap: those defences are instruction-level, not yet structurally enforced (agent sandboxing and audit logging still to build).

A full governance operating system in daily use; an enterprise CRM transformation across hundreds of users; reusable measurement frameworks.

20 years of websites; a live Next.js directory; a daily-use finance app. The modern typed-React stack is newer than the two-decade web foundation (honest).

Launched a 16-unit complex at 22% net profit; a startup proposition to £1.2m in 18 months; scope decisions made and documented, not drifted into.

CI/CD live across four repositories with auto-deploy on push, plus Vercel deployments and 20 years of domain/DNS/hosting. Honest caveat retained: no infrastructure-as-code or formal monitoring yet.

A reusable, themeable engine that renders on-brand explainer videos (authored HTML → headless render → MP4), now spanning two brands; plus generative-model work — an AI-built artist concept (Suno music + AI imagery). Honest caveat: the video path is code-rendered, not model-generated, and the generative work is one concept, not yet a standing operation.

First external consultancy deliverable shipped; 39 skills in daily use. Honest note: the repeatable consultancy pipeline past one client is still to be built.

A 25-year marketing career at scale (HSBC £50m+, Santander). The AI-automation layer specifically is still developing — the level reflects that, not the career depth.

Two complete consent and disclosure regimes designed and live on two production sites — cookie consent gating, privacy/cookies/terms, AI-use disclosure — each guarded by a self-test that fails against a deliberately broken implementation, plus a primary-source regulatory correction sweep in which two published claims were downgraded rather than left unsupported. Honest note: self-assessed UK GDPR/PECR implementation, not a legally reviewed programme.

A strong enterprise commercial and transformation track record; the AI-specific enterprise overlay is pilot-stage, not yet org-scale (honest).