AI Design Blueprint is a public governance standard and validation layer for agentic AI. I turned ten principles of supervised delegation into a public MCP server, an executable validator and replayable readiness reports. Its own bridge self-audit improved from <strong>35/F to 100/A over 14 iterations</strong>, while failures and limitations remained visible.

I replaced anchored consensus with unanchored consensus because the evaluator had been reading its own prior verdict, which made stable scores misleading.

I accepted extra evaluation cost to avoid false stability, using unanchored runs even though they required more work than anchored consensus.

I made only production-blocking findings lower the headline grade, while keeping hardening and polish visible so readiness did not require perfection.

I required an audit row before granting a production-ready badge and failed closed on write errors, giving up availability to preserve inspectability.

I kept submitted code out of persistence except for an opt-in certification hold, accepting that the validator could not fully distinguish regressions from pre-existing misses.

Improved the published bridge self-audit from failing readiness to certified top-grade readiness over 14 iterations.

Achieved top ratings for license, quality and maintenance in the Glama registry.

Converted anonymised closed-beta tester feedback into shipped product fixes.

A practical standard for governing agentic AI.

AI Design Blueprint

The 2015 BBC iPlayer TV Guide redesign, led as Senior UX & Interaction Designer in London. The goal: make live television fast and obvious to find inside a dense, multi-channel schedule, and rethink how viewers open a live programme, without losing the density that makes cross-channel browsing useful.

I put the pink On Now line and a persistent route back to now ahead of other refinements because live was what viewers came for and deserved priority.

I organized the guide around viewer intent rather than internal content structures, so Search, Browse, Watch and Manage could offer clear routes that converged on the same content page.

I added a consistent route back to On Now, assuming it would keep live one tap away; later testing confirmed that it fixed the difficulty of returning after scrolling.

I replaced Yesterday, Tomorrow and Weekend with day parts because testing showed those links were less useful than parts of the day.

I used progressive disclosure to keep the guide scannable, accepting that viewers saw only a title and time until they focused a row to reveal more detail.

Reduced the time users needed to find live programmes, improving access to live content.

Increased engagement with the redesigned TV Guide experience.

Improved overall user satisfaction following the redesign.

Find live programmes across BBC and other channels.

BBC iPlayer TV Guide

I redesigned BBC iPlayer TV Home to combine editorial discovery with a channel-by-channel view of what was on now and next. As Senior User Experience Designer, I shaped the responsive experience across desktop, tablet and mobile, from research and wireframes through prototyping and testing.

I placed the prime-time carousel above the schedule so editorial discovery and live utility could coexist on one page, rather than competing across separate destinations.

I preserved the distinction between choosing something now and planning what to watch next across breakpoints, rather than compressing the desktop layout, because viewers needed those decisions to remain clear.

I used tabs for the mobile schedule instead of compressing it into one view, retaining the same information architecture in a narrower viewport. This kept the schedule's distinction between current viewing and planning ahead.

I postponed the stronger visual representation with content posters to a follow-up iteration because it required development effort in the short term, delaying an option users appreciated.

I separated ON NOW from ON NEXT because I expected viewers to distinguish immediate viewing from planning ahead, and later metrics showed the states were easy to use and used for switching.

I deferred the stronger poster representation to a future iteration, giving up a user-appreciated visual treatment in the initial release while short-term development effort remained a concern.

The redesigned TV Home experience shipped across desktop, tablet-width, and mobile layouts.

Making live TV easier to discover, choose and plan across every screen.

BBC TV Home

I led UX, UI, interaction design, prototyping and bi-weekly usability testing for the BBC iPlayer Downloads app on Mac and PC. The rebuild introduced Series Record, episode grouping, watched progress and clearer offline playback and expiry states.

I grouped episodes under a series because that made My Downloads reflect how viewers thought about programmes instead of how files were stored. Counts for unwatched episodes, watched progress and availability helped people choose what to play without opening every item.

I made automatic downloading a user-controlled setting instead of a hidden default. I kept Series Record visible and explained what would happen to future episodes so viewers could understand the commitment.

I made expiry an explicit, understandable state rather than silently deleting content, accepting a short period of visibility before automatic removal so viewers could see what had expired.

I made expiry an explicit state and kept expired episodes visible briefly before automatic removal. I accepted the short period of visibility people wanted, even though it delayed clearing expired episodes from the library.

Audience survey demand for the Series Record feature, which automatically downloads new episodes of a subscribed series.

Making offline viewing more dependable through Series Record, episode grouping and visible playback states.

BBC iPlayer Download Manager

The 2020 Direct Ferries post-booking app redesign, led as Lead Product Designer. After booking, ferry customers often don't know what to do, where to go, or whether their ticket will work at the port. I built operator-personalised, task-driven guidance (what to do, when, and where) to prevent missed sailings and cut support contacts.

I replaced the information-heavy booking page with a task-driven layout because operator differences were easy to miss and the next action was unclear.

I used rules-based content templates instead of hand-building a screen for every operator, because one shared structure could scale across uneven APIs.

I kept the navigation within a stable Direct Ferries structure and limited operator-specific rules to moments when they changed what customers needed to do, giving up full operator white-labelling across every screen.

I treated one component set as able to take operator-specific rules and data, then watched support tags, engagement and failure rates before widening coverage.

Customer NPS improvement in customer satisfaction

Support contact reduction change in operational efficiency

Time efficiency reduction in operational efficiency

The quickest way to view, use and manage your booking

Post Booking Experience App

I led the product side of the loveholidays rebrand from 2023 to 2025, turning a new brand direction into shippable product across the highest-value booking journeys. The work had to land without hurting conversion at checkout, while removing the price ambiguity and buried reassurance that were making customers hesitate. It shipped in phases ordered by commercial risk, each one carrying agreed guardrails and a rollback trigger.

I chose a dual-theme design-system rollout over a single big-bang UI rewrite. I made that choice so funnel-critical journeys could migrate incrementally while the conversion and performance guardrails stayed armed.

I regrouped SRP filters into high-intent clusters and used progressive disclosure, giving up immediate prominence for rarely used filters to reduce cognitive load and protect performance on a high-traffic page.

I prioritised price and inclusion emphasis at responsive breakpoints, giving up pixel-perfect alignment so layouts could tolerate longer localised copy without breaking the hierarchy.

Reduced contacts per booking, with the largest decreases tied to pricing, inclusions, cancellation and change issues.

Improved click-through from search results pages to product detail pages after rebalancing the information hierarchy and filters.

Improved page loading performance through font subsetting and preloads, validated with real-user and lab performance testing.

Localisation became real during expansion, including the Germany launch as weloveholidays on 2 May 2023 , and brand expression deepened further with campaign activity such as the first national brand campaign on 1 Feb 2024 .

Customers more successfully understood baggage, transfers and board basis on the product detail page.

Customers better understood cancellation fees and expected refunds in Manage Booking.

I led the product side of the loveholidays rebrand from 2023 to 2025, turning a new brand direction into shippable product across the highest-value booking journeys. The work had to land without hurting conversion at checkout, while removing the price ambiguity and buried reassurance that were making customers hesitate.

Shipping a new brand through the booking funnel without giving up conversion.

loveholidays Rebranding

A design-led, production-ready internal AI platform that made customer and business data explorable through natural-language conversation. Built for eight departments at loveholidays, it reduced analysis from weeks to seconds, cut manual analysis effort by more than two-thirds, and supported a shift from reactive reporting to proactive decision-making.

I chose a living layer for asking different questions directly of the customer journey over another static diagram or dashboard variant, because teams needed to explore customer questions instead of relying on a fixed view.

I treated exploration before answers as the premise, then checked it by having analysts and department leads test live natural-language queries in the prototype and beta.

I represented journeys, segments, markets and feedback through a semantic abstraction layer instead of exposing underlying systems directly, accepting less direct access in exchange for adding models, tools and data sources without disrupting the conversation pattern people had learned.

Reduced the time required for manual analysis during the internal beta.

Enabled exploratory customer and business data engagement across multiple departments during the beta.

A design-led, production-ready AI platform that democratised data, accelerated discovery, and helped us move from reactive reporting to proactive, AI-powered decision-making.

Internal Multi-Agent AI Data Platform

A trust-led Hotel Details Page rebuild for loveholidays, combining verified customer reviews, clearer room comparison and modular experimentation to help customers resolve booking doubts before payment.

I chose independently testable blocks instead of waiting for a single reveal because one release would have concentrated data, content, tracking and delivery risk. I accepted repeated testing so I could measure and roll back each change independently.

I kept TripAdvisor alongside our own reviews instead of replacing it because customers already understood it as a trust source. The combined sources gave the page independent proof and verified evidence from customers who had booked and travelled with loveholidays.

I accepted repeated test overhead and less visual unity at the beginning in exchange for independently testing each block. That let me measure, learn from and roll back changes without losing the others.

I shipped reviews before rooms because reviews were the highest-confidence, lowest-dependency quick win. I waited for the data-quality work needed to make room comparison trustworthy.

I kept both review sources and accepted the additional content and maintenance responsibility. I did so because they gave the page independent proof alongside verified evidence from customers who had booked and travelled with loveholidays.

Loveholidays launched its own integrated customer reviews as a public market milestone.

Directional increase in booking conversion on the measured page.

Rebuilding loveholidays’ hotel page around verified customer proof, clearer room decisions and expectation-setting before booking.

Hotel Details Page

MatchBoss was Squawka's free-to-enter, real-money football prediction product. I designed the consent, selection, submission, scoring and live-match experience across the Euro 2016 beta and the 2016 to 2017 Premier League season.

I trusted the funnel over the stated preference and retitled the screen rather than shortening step three. The funnel identified it as the worst screen, while the survey said the form was fine.

I rewarded failed actions instead of keeping a simple realism rule, because every pick could still move in the live view.

I rewarded attempts as well as outcomes to keep picks meaningful during the match, giving up a simple realism rule for a live view where every pick could still move.

I used three tiles and three pages to explain the prize structure rather than turning the landing page into a comprehensive terms document. Its first job was getting someone to a first pick.

I shared the submission rather than the result, which supported acquisition but gave up a return prompt when users finally had something worth boasting about.

Average open rate for in-play emails across gameweeks 6 to 13.

Average click rate for in-play emails, the highest click rate among the emails sent.

Open rate for results emails, which was the highest email open rate.

Sign-up conversion improvement in conversion performance

Order conversion change in conversion performance

A full-season football prediction product where the game was easy to play, but difficult to return to.

Squawka MatchBoss

A painter's fifty-year catalogue rebuilt as a public archive that visitors explore through a curator agent, and that the artist maintains through a private studio agent by conversation. The migration preserved 314 items and every old address, 1,222 checked live on 4 October 2026, while adding clearly labelled draft texts and the artist's approval for every public change.

I chose an approval card and typed confirmation over one-tap publishing because a proposed correction was not the same as an approved change, and Gaetano needed to decide what reached the archive.

I chose retrieval before writing stories because chapter-by-chapter generation failed in 9 of 10 runs, while retrieval-first failed in none of 16.

I accepted extra friction from the approval card and typed confirmation because a proposed correction was not the same as an approved change to an artist’s public record.

I relied on the matcher to link candidate material to catalogue works, then checked its results against hand-matched works to see whether its suggestions held up.

Preserved and checked the full catalogue during the archive migration.

Added draft texts for works that previously had no text, helping fill catalogue gaps.

Created the final set of missing-work texts at a total cost of $8.45.

Verified old addresses still lead to the correct page or picture, preserving access to the archive.

A painter’s fifty-year catalogue, rebuilt as an artist archive visitors can explore and its artist can maintain.

Gaetano Zampogna Artist Archive

I led the responsive rebuild of The Savile Row Company’s Magento storefront, covering information architecture, wireframing, responsive UX and UI, CMS merchandising modules, front-end implementation and Magento delivery across desktop, tablet and mobile.

I chose a shared responsive template system over a separate mobile site because it kept navigation, trust messaging and cross-selling consistent while letting each layout adapt to available space.

I chose reusable CMS banner variants instead of independent seasonal placements, accepting a more constrained component system so merchandising could publish campaigns repeatedly across the site.

I added a product-detail cross-link and quick-view popover because I believed in-context information was more relevant at that stage, then used A/B tests and customer surveys to check that belief.

I defined the grid and information architecture before drawing the full templates, which gave the storefront shared structure rather than a collection of individually optimised pages.

A single responsive Magento storefront operated across desktop and mobile screen sizes, replacing the desktop-only build.

Documented breakpoint rules enabled Magento templates to scale predictably from desktop to mobile.

The site structure was mapped across six core information pillars before template design.

Documented a five-breakpoint responsive grid spanning desktop through mobile layouts.

Rebuilding a desktop-only Magento storefront into a responsive luxury menswear experience.

The Savile Row Company

Virgin Media LineUp Builder was a Squawka-delivered Euro 2016 campaign tool. I designed a guided three-step lineup builder using live Opta data, teaching the mobile gesture before asking fans to create and share a team.

I chose formation before players instead of an open canvas. I fixed the available slots, removed invalid states and kept the pitch legible on a small screen, so fans never faced a blank pitch beside 23 names.

I chose live data over a static list because the squad-announcement timetable forced the decision and the credibility benefit justified it.

I anchored the tool to Euro 2016 to gain timeliness and cultural relevance. I accepted a shorter product life beyond the group stage and less meaning outside Europe.

I placed an email request at completion to serve the commercial ask. I gave up an easy sharing moment when fans most wanted to post and least wanted a form.

I put a four-step tutorial before the first pick because I believed fans needed the tap gesture demonstrated on a phone during a tournament. I checked the flow with beta testing among football fans.

User engagement rate change in user engagement

A guided, mobile-first lineup builder that gave Virgin Media a way into the Euro 2016 conversation without tournament rights.

Virgin LineUp Builder

As Senior Mobile UI Designer, I designed Watsons mobile commerce concepts and an implementation-ready style guide for regional teams. The system covered campaign discovery, product browsing, filters, product detail, store finding, localization and reusable UI patterns across multiple Asian markets.

I chose a persistent utility bar for Menu, Member, Store locator and My List so shoppers could reach core actions one tap away while campaigns rotated above.

I assumed the system could roll out across regional business units, and I checked that assumption with a three-screen concept showing the same structure in Chinese and Thai.

Style-guide system referenced four Asian markets for regional implementation.

The concept demonstrated localized Chinese and Thai market interfaces.

Designing a localisable mobile commerce system for campaign discovery, product browsing and store finding across Asian markets.

Watsons Mobile Commerce System

Lead Product Designer on the what3words map app, 2018. The company had a working address system and very few people using it, and the research found that the map was the barrier rather than the addresses. This is what we changed and what it cost.

I showed the grid only at zoom 19 because city-level squares shrank below a pixel and obscured the map, accepting that new users would need to zoom far before seeing the product idea.

I used separate marker shapes for floating, locked, saved and named locations because distinct visual states let people read the map at a glance while the location was still moving.

I made address type resize to preserve all three words across languages, rather than use a fixed size and cut text, because removing any word would make the square unfindable.

I built on Google Maps to keep the map familiar and rely on a proven integration, giving up control of the underlying cartography and putting the visual difference in the layer above.

I accepted that first-time users would not see the product idea until they zoomed far in, because showing the grid earlier would turn tiny squares into a grey wash over the map.

I bet that users could use the grid and satellite view at high zoom to visually find a familiar, preferred position. I checked that idea with a focus group and recordings of live production.

Languages the three word address system was published in.

Average three word address length across the top 10,000 addresses per language, lowest (Chinese) to highest (French).

Longest three word address per language, lowest to highest.

Helps everyone talk about everywhere

what3words Map App

I led the UX and product design for what3words Map Web, a responsive browser product for finding, refining, saving, sharing and navigating to precise 3m × 3m locations through a map-based experience.

I built on Google Maps to give people familiar map behaviour, accepting the risk that what3words could appear to be another map or navigation tool.

I kept the selected address locked while moving from standard road view to satellite imagery, because that let people confirm the general area and exact spot without losing the address they were validating.

I kept search, refinement, and saved locations in one map-anchored Inspector instead of separate destinations because the Mark could evolve as a person moved from an unsaved location to a saved and then labelled location.

Share of MVP usability-testing participants who completed saving a location.

Share of participants who found the location-saving flow not difficult.

Share of participants who understood what what3words was after using the product.

Designing a responsive map product that helps people find, refine, save, share and navigate to any precise 3m × 3m location.

what3words Map Web

3WordPhoto extended what3words’ 3-metre addressing system into photography, allowing users to capture or import photos, refine their location, and embed a customizable three-word address label. The app supported personalized editing and sharing through social media, messaging apps, and links to precise what3words locations.

I replaced the fixed watermark with a movable, resizable label because corners often covered the subject. I accepted an extra placement step on every photo so photographers could put the label where it worked.

I added a map panel for manual pin refinement rather than treating GPS as final, because GPS could land in the wrong three-word square. I accepted an extra screen to let photographers choose the precise spot.

I opened the camera at 4:3 and animated the crop when photographers chose another ratio, rather than asking for a ratio up front. That showed framing on the live image instead of explaining it in a settings list.

I blurred the image beneath the camera tools and address panel to keep the photo visible while it was composed. I accepted higher rendering cost and added contrast treatment for the label because the blurred background could vary.

I used a standalone photo app to test whether download interest could justify adding the feature to the main map app, treating download counts as a signal before that integration decision was proven.

Embedded precise location addressing based on three-metre by three-metre squares.

The app's language support broadened access for users across international markets.

Making the place a photo was taken visible in the photo itself.

3wordPhoto App

An OCR-based mobile app designed for Mongol Post to scan what3words addresses, sort parcels into delivery-area bins, and calculate optimized delivery routes. The product addressed Mongolia's lack of consistent street addressing and vast, sparsely populated geography, improving sorting accuracy, delivery success, and operational efficiency.

I bet that OCR could connect imperfect physical labels to a bin and route while preserving a fast human workflow, and I treated explicit failure, validation, confirmation, and recovery states as essential to that outcome.

I made failed OCR reads visible and provided a recovery path instead of silently advancing workers, because errors needed to surface before the next item entered the workflow.

I chose a simple, large letter that appeared live during scanning instead of visualizing the bucket card, because I wanted the simplest possible bin cue to appear during the scan.

I provided a way to open the native device map for full geolocation context, accepting that traffic information and other map content would not be embedded in the app.

Increased the share of deliveries completed successfully on the first attempt.

Improved parcel sorting accuracy in the delivered OCR app and partnership results.

Reduced the time required to process each delivery.

Sorts and routes deliveries by scanning a 3 word address, even offline.

what3words OCR Sorting & Delivery App

The project tested whether voice commands could provide a compelling and reliable way to request transportation, while demonstrating the commercial potential of integrating what3words with established ride-hailing services. Research and testing focused on voice-recognition accuracy, user trust, conversational interaction and seamless third-party API integration. The released concept received positive feedback from users, investors and partners for making taxi booking and destination entry more convenient.

I used brightness and contrast changes in the tiled grid to signal listening or speaking instead of a conventional microphone icon or explanatory copy, keeping the interaction calm and glanceable in an automotive context.

I sent a “no” response back into correction rather than ending the journey, because a misheard three-word address needed a recovery path before it could trigger an incorrect booking.

I treated voice commands as a candidate for compelling, reliable taxi requests and tested recognition accuracy, user trust, conversational interaction, and third-party integration to see whether that premise held up.

I put more time into discovery and research, accepting the upfront effort of testing with existing map-app users so I could gather feedback before using the proof of technology for partner acquisition.

what3words resolved every 3m x 3m square worldwide, enabling precise voice-friendly destination entry for taxi booking.

The work contributed to integrations across automotive products and smart-assistant platforms.

The project tested whether voice commands could provide a compelling and reliable way to request transportation, while demonstrating the commercial potential of integrating what3words with established ride-hailing services. Research and testing focused on voice-recognition accuracy, user trust, conversational interaction and seamless third-party API integration.

Navigate to a 3 word address by just speaking it.

3wordTaxi App

skills

Define problem framing, success metrics, and AI-readiness (data, privacy, safety). Align stakeholders on scope, value hypothesis, and guardrails.

Concept Development

Low-fidelity flows for core tasks, agent handoffs, and conversational turns. Validate information architecture and decision points early.

Wireframing

High-fidelity interactive prototypes (web/mobile/voice) including LLM prompts, memory states, fallbacks, and safety/consent patterns.

Prototyping & Mockups

Experiment on copy, layout, and AI behaviours simultaneously; monitor guardrail breaches and quality KPIs alongside conversion metrics.

MVT (Multivariate Testing)

Run controlled tests across journeys and agent interventions; measure task success, latency tolerance, and perceived trust/safety.

A/B User Testing

Design for accessibility, performance, and device ergonomics; ensure conversational and multimodal UI scale across breakpoints.

Responsive Design

methodologies

End-to-end journey design, including agentic flows, error recovery, and measurable outcomes tied to product strategy.

User Experience Design

Design systems, tokens, and patterns for scalable UI; stateful components for AI responses, loading, and uncertainty.

Explore the Case Studies