AI has changed how we make products. It hasn't changed who we make them for.
Google / Exec Circle
Context
Google UK ran Exec Circle through the first half of 2021: a weekly executive webinar series in collaboration with Stanford Graduate School of Business, connecting Google's key business partners with Stanford academics during COVID. I was sole product designer and front-end engineer at Cheerful Twentyfirst, designing and building the platform end-to-end.
Problem
The platform needed to support a weekly programme for busy executives who wouldn't return if the experience didn't justify the time. Pre-event content had to drive attendance; post-event content had to keep the conversation going between sessions. And it had to carry Google's brand without looking like a generic webinar tool.
What I did
A content-rich event hub: calendar, resources library, speakers and attendees directories, session cards with Stanford co-branding. The resources page (documents, videos, podcasts per session) became the engagement surface. I designed every surface and built the front end.
Outcome
- 72% of invited executives accessed the platform; average 7+ resources each
- 4.9/5 overall series rating
- Google commissioned new event series based on Exec Circle
Thank you, Google. Exec Circle has been very enlightening, inspiring and great for building a peer network.
Chief Marketing Officer, New Look
Mercedes-Benz / Media Decision Engine
Context
Mercedes-Benz's Media Decision Engine is a paid-media analytics platform used by planners to track, measure, and optimise advertising performance. Built by OMD (Omnicom's media agency), it had grown over years without dedicated design input. I was lead product designer, working with a UX researcher and the platform's original builders at OMD.
Problem
Users gave up within ten minutes. The platform had roughly 15 tools, each requiring its own configuration: market, vehicle model, competitor, fuel type, timeframe. A planner running a single scenario had to re-enter the same parameters in every tool. Two to three people used it daily.
What I did
- Information architecture. Regrouped the tools into three process stages, ordered by how planners actually work.
- Global scenario settings. Consolidated shared parameters into one overlay reachable from anywhere, each setting entered once. Local parameters stayed in their tools, clearly labelled.
- Onboarding. First-use guided tour, expert-pitched copy, empty states that guide rather than block.
Outcome
Within two weeks of test-market release, average session time rose from under 10 minutes to over 30 minutes. Daily active users grew from 2-3 to 40-50 across the account. (James Hutchinson, OMD Data Operations Manager.)
GSK / Core Design System
Context
GSK's 2022 rebrand by Wolff Olins created a bold visual identity. It worked for consumer marketing. It didn't work for GSK's digital products: the primary orange failed WCAG AA in some contexts, the consumer-marketing tone didn't suit specialist tools for healthcare professionals, and no shared design system existed.
What I built
GSK's first company-wide design system:
- 500+ tokens, accessibility resolved at the token layer so teams get compliant colours automatically
- A density mode for specialist tools that need more information in less space
- iOS and Android extensions. WCAG 2.1 AA throughout
Adoption through a cross-team working group I ran in weekly workshops. The five key product teams, the groups that create most of GSK's digital products, adopted the system. In production on gsk.com and Brandhub (GSK's digital asset management platform, also mine).
The arc
The design system was the culmination of Interbrand's GSK account, not its foundation. I started with AUREA (an antimicrobial-resistance data tool for healthcare professionals), then GSK Medical, then GSK Direct (B2B vaccine ordering). Each engagement was won by the previous one's success.
Outcome
Five products shipped. The system is company-wide.
Tradu / Jefferies
Context
Tradu is a multi-asset trading platform, wholly owned by Jefferies Financial Group. Built with raw colour values, no design tokens. The brief was one week: design a light mode for onboarding.
What I did
The existing dark mode used hardcoded colour values. I worked backwards: defined primitive tokens from the values in use, then derived semantic tokens that both modes resolve to. One abstraction, two modes.
The one-week brief expanded across three statements of work:
- Light mode token retrofit
- Onboarding UX audit: 19-screen flow, reducing taps, rethinking progress indicators, redesigning field states
- Corporate site UX audit
The strongest decision was on accessibility. Tradu's brand guidelines specified a darker shade of their accent orange for light mode. The standard shade was lighter, higher contrast, and passed WCAG AA more cleanly. I recommended the standard shade universally. Tradu amended their brand guidelines.
Outcome
~45s and 11 taps removed from the onboarding flow (measured). One-week brief earned three engagements.
WSJ / CEO Council
Context
The Wall Street Journal's CEO Council is one of their most prestigious events: 250+ CEOs and global leaders, traditionally hosted in person. COVID made that impossible. Cheerful Twentyfirst, who had run the in-person event, was asked to deliver the virtual edition. Eight weeks to design and ship.
Problem
A time-poor audience with no tolerance for friction. Networking was the reason people attended in person, and virtual platforms didn't support it well. And it had to carry WSJ's brand.
What I did
- Two spaces: Auditorium for talks and Q&A, Lobby for networking. Physical metaphors that made navigation intuitive without explanation.
- Page fold as priority: above the fold on every page, the single thing the user most likely wanted (question form in the Auditorium, live attendee count in the Lobby).
- Built for reuse: utility classes, CSS variables as design tokens, documented as we wrote it.
I designed every surface and led front-end development alongside Gramercy Tech.
Outcome
2.4x attendee engagement versus WSJ's prior virtual event (Cheerful Twentyfirst). The design and codebase were adopted across WSJ's Leadership Institute events: CFO & COO Council, Technology Council.
A huge thank you for the delivery. We are getting endless thanks, high-fives and praise from all parties, internal and external.
Angus Peckham-Cooper, Vice President, Global Events, Wall Street Journal
I was making this case in 2023. Here's what I'm building today:
Tenon / Roles that fit
Roles that fit
Tenon finds the roles that fit you, then helps you earn them with the right moves and materials.
Problem
AI made it easy to apply to everything, so people did. Interview rates per application have dropped from 15% to 3% (CareerPlug, 2025). 67% of hiring managers say AI-generated applications have slowed the hiring process (Robert Half, 2026). 70% of hiring managers say AI helps them make faster and better decisions (Greenhouse, 2025). 8% of candidates believe it makes hiring more fair (Greenhouse, 2025).
Product
I am the user: Tenon grew out of my own job search. It builds its understanding of who you are, shapes an approach for each role, and generates the materials to back it up. Real agent calls, a purpose-built design system, structured outputs: designed around the hard problems in agent-driven apps, consistency and trust. Months of product work in weeks, building with AI. I use it for every role I pursue.
Design system
The design system has a fixed set of building blocks: value displays, timelines, input controls. The agent picks which blocks to use and what content goes in them, but it can't change how they look or invent new ones. Each block validates its content before it renders.
The reasoning: if the agent controls appearance, every prompt change risks breaking the look. If the component enforces its own appearance, the agent composes freely without breaking consistency. Appearance decisions are made once, in the component, and held.
This pattern shipped as open infrastructure through 2025-26: Google's A2UI and Vercel's json-render landed on the same architecture: the agent fills pre-built components from a fixed set, and anything not in the set doesn't render. Tenon arrived there independently, with a smaller, more opinionated vocabulary.
Trust
Tenon shows the basis behind every recommendation, not a confidence score. LLM confidence is systematically miscalibrated (Xiong et al., 2025), and showing confidence percentages actively backfires: in controlled tests, uncertainty highlighting was the worst of three interventions for helping users calibrate their reliance (CHI 2025). Instead, each recommendation surfaces what it's based on: which parts of your profile matched, which signals from the role informed the assessment, and where the evidence is thin.
The more a decision matters, the more Tenon checks with you. New content flows without interruption. Content you've already seen or decided on shows old next to new so you can compare. Actions that change your state (dismissing a role, submitting materials) get a clear notice. This avoids the two failure modes: asking permission for everything, and asking permission for nothing.
Generation
Each role gets a composed approach: why it fits, what angle to take. Tenon then generates the materials to back it up. Blocks appear as they're generated, with honest placeholders where the agent is still working.
The underlying pattern: layout, structure, and vocabulary are fixed. The content is composed per role, tailored to your positioning and the specific opportunity. The frame is reliable. The content is specific.
What this is
A working product, designed and built end-to-end: research, product definition, design system, front-end, agent integration. Consistency, trust, and composition quality in agent-driven interfaces are design problems. I built a product to solve them.
Back from a year travelling. In my free time I keep up with F1 and overanalyse media. I'm counting the days until Control Resonant is released and I can spend some quality time with Ahti the janitor.
Summer '25 bike adventure