01
Visual source of truth
I rebuilt MedExpress rebrand pages as a 60-component HTML/CSS reference so the team and the tools could work from the same inspected interface.
Hey, what would you like to know about Andy's work?
The static page is available, but the Clyde API is not reachable from this view. Run npm run dev locally or deploy the API functions to unlock the prompt.
Scroll down to find out more...
Principal-level AI product architecture
I design AI products by making the system visible.
I design the layer around AI that makes it usable: source architecture, interfaces, model boundaries, and evaluation loops. The work here is recent, shipped, and inspectable: Clyde, a MedExpress design memory, and a regulated competitor-intelligence product.
- Shipped systemsLive RAG assistant, team design memory, prototype gates, and competitor-intelligence tooling
- Hands-on AI fluencyModel routing, prompt contracts, design-to-code workflows, retrieval, and evaluation
- Craft with constraintsInterface decisions shaped by evidence, brand quality, clinical safety, and operational trust
Live product work
The story starts with the tools themselves.
Clyde sits inside the portfolio because the work is not just the write-up. It shows the same decisions in use: what the system can see, how answers are grounded, what the interface makes explicit, and where the boundaries sit.
Clyde
Live portfolio assistant with citations, access controls, and server-side Gemini responses
Leo
Shared MedExpress design memory packaged for the team that used it
Crawler
Gemini and Playwright workflow for regulated consultation journeys
Gate
Prototype review workflow with automated checks, specialist review, and clear approval states
On-screen craft
The work also has to feel clear and trustworthy in use.
The architecture matters, but so do hierarchy, pacing, states, and restraint. These artifacts show how that work lands in something a team can review, use, and trust.
02
Inspectable intelligence UI
Competitor findings are presented as reviewable product evidence, with persona context, criteria, screenshots, and decision cues rather than a loose deck.
03
Clyde in use
The assistant surface shows the product decisions directly: grounded answers, visible citations, and clear answer boundaries rather than hidden orchestration.
04
Team design memory
The Leo bundle captures the reusable operating memory behind the work: source order, answer rules, maintenance constraints, and prompts the team could share.
Shipped outcomes
The same approach has already shipped at meaningful scale.
These are concrete scale signals, not inflated business claims. They show the shape of the work: clearer ground truth, faster iteration, and AI systems teams could actually use with confidence.
£805M
annual MedExpress UK funnel represented in the design, analytics, and acquisition system work
9
evidence-sourced GLP-1 personas used to pressure-test journeys, content, and prototype directions
22
UK healthcare competitors modelled in the crawler, dashboard, and pricing intelligence surface
41
iterations of the shared MedExpress design-agent bundle shipped as team-facing AI infrastructure
Case studies
The case studies show how those systems were built.
Output / RAG system, Turnstile protection, Upstash Redis API quotas
Designing Clyde: The grounded AI portfolio assistant
I designed and built Clyde, a principal-level AI assistant that demonstrates model-in-the-loop interfaces, BM25 and vector retrieval fusion, quota guardrails, and server-side safety boundaries.
- My role: interaction design, retrieval architecture, Turnstile integration, prompt engineering, and security proxy.
- What shipped: live grounded RAG assistant, reciprocal rank fusion pipeline, password unlock state, and Vercel proxy.
- Why it mattered: visitors can inspect the prompt, retrieved sources, and refusal boundaries directly in the interface.
Output / Design memory, Leo bundle, prototype gate
Designing the MedExpress AI design memory and acquisition workflow
I owned the product architecture for a system that turned scattered brand, clinical, analytics, research, and component knowledge into reusable design infrastructure for humans and AI agents.
- My role: knowledge architecture, source hierarchy, assistant contract, quality bar, and prototype workflow.
- What shipped: 212-file knowledge base, 60-component rebrand registry, 41 Leo bundle iterations, and quality gate.
- Why it mattered: new design work started from product truth rather than Slack archaeology or model improvisation.
Output / Intelligence dashboard, model boundaries, crawler system
Building AI competitor intelligence for regulated medical-intake flows
I designed the system and operator surface for repeatable competitor intelligence: personas, page hashes, answer maps, Gemini form-solving, route persistence, ethical stops, and dashboard review.
- My role: problem framing, architecture, persona model, dashboard UX, anti-bot policy, and model boundaries.
- What shipped: 22 competitor targets, 16 personas, 57 protected test files, and pricing intelligence surface.
- Why it mattered: competitor knowledge became refreshable, queryable, and inspectable instead of quarter-old screenshots.
Practice
This is how I turn AI into product infrastructure.
I am most useful where the work is ambiguous, technical, and quality-sensitive. The model is only one piece; the real design problem is building the system a team can use, inspect, and trust.
01
AI product architecture
I shape source models, retrieval paths, model roles, guardrails, and the experience people actually use.
02
Evidence systems
I turn research, analytics, compliance, and stakeholder decisions into durable source systems for teams and agents.
03
Hands-on building
I work directly in code, prototypes, RAG pipelines, prompt contracts, dashboards, and design-to-code workflows.
04
Craft under constraint
I keep visual quality, interaction clarity, clinical safeguarding, accessibility, and trust in the same decision loop.
About
I do my best work in ambiguous, regulated, evidence-heavy products.
I design AI-enabled products by connecting customer needs, business pressure, source quality, technical constraints, and the experience that has to carry the decision. My recent work has focused on MedExpress UK and HeliosX, across acquisition, consultation, checkout, clinical review, retention, design-system memory, RAG assistants, and competitor intelligence.
The thread through the work is simple: clarify the messy problem before the interface appears, build an experience that makes the decision inspectable, and leave behind a system that lets the next team start from a higher floor.