The problem

When I joined Weed.de, the patient journey was scattered across multiple disconnected steps. Patients had to manually search for a doctor, explain which products they wanted, wait for a consultation, upload their prescription separately, and then find a pharmacy on their own. There was no guided flow tying any of it together, which led to confusion, high drop off, and only around 20,000 orders in the entire first year. Making things harder, medical cannabis still carries real stigma, even where it is fully legal. Patients were often skeptical, having encountered grey market operations before, so anything that felt unclear or unpolished would break trust immediately. Every design decision also had to navigate German healthcare regulations alongside SEO requirements, balancing what users needed to understand, what the law required, and what search engines needed to index.

Redesigning the patient journey

I redesigned the entire patient journey into a step by step wizard completable in under ten minutes. Patients could browse products, check real time pharmacy availability, complete a short medical questionnaire, choose a doctor, verify their identity, and sign documents, all without leaving the platform. Prescriptions were automatically routed to the selected pharmacy, and patients were notified the same day when their order was ready. Each step collects only what’s needed before moving forward, so patients never face a wall of forms or unclear requirements. The nine step structure breaks a traditionally intimidating medical process into manageable decisions, with persistent progress indicators reducing drop-off from uncertainty about how much is left.

Balancing UX, compliance, and SEO

Designing in a regulated healthcare market means every design decision has a legal dimension. I worked closely with the legal and SEO teams to find design patterns that satisfied all three requirements without creating a fragmented experience. Products could not display effects, and we weren’t allowed to directly showcase images of cannabis flowers. Constraints like these shaped how every product page was designed. The solution was a layered content architecture: a clear, scannable summary for the patient on top, detailed specifications and legal disclosures below, and semantic structure and metadata woven throughout for SEO.

Mobile-first approach

More than 80% of our users came from mobile. This was the starting point for every design decision. I built a layout system that started from the mobile experience and scaled up to desktop without sacrificing usability at either end. Every flow was conceived at the smallest breakpoint first. Components were built to reflow and restack naturally, so a process screen or product detail page on mobile would expand into a desktop layout with minimal additional design work. This had a direct impact on development speed. Because the mobile layout was the source of truth, scaling up to tablet and desktop became a matter of adjusting container widths and repositioning elements within an already proven information hierarchy.

Where the paper went

The Comitetul Național Paralimpic looks after about 55 clubs, 189 active athletes and some 30 coaches. Every part of that relationship lived somewhere different. A club applied for affiliation with a form and a stack of scanned documents. Athletes were registered, moved between clubs and entered into competitions by email. A head coach asked for a travel budget with a referat de necesitate, a document that went from desk to desk collecting signatures. Fees were checked by hand against bank statements.

The brief was one platform for four roles: club administrators, CNP’s operators, CNP’s admins and the head coach of each sport. Athletes are in it as people, not as users; they never log in. It had to be in Romanian, meet WCAG 2.1 and keep its data in the EU under GDPR. And it had to be simpler than paper for people who had used paper for years.

Listening before drawing

I didn’t start in Figma. Every call with CNP’s team was recorded and transcribed, and agents turned each transcript into the project’s knowledge base the same day: one page per call, the decisions made in it, the questions it left open. By the time I drew a screen, I could trace it back to who had asked for it, and when.

I wanted to hear from the people on the other side of the paper too. I sat down with a club administrator and went through the folder their club sent to CNP every year, page by page. I talked with head coaches about the referat and where it got stuck, and with athletes, reached through their clubs, about the moments the paperwork landed on them. Those conversations were recorded as well, and the agents folded them into the same research wiki the whole team read.

The calls changed the brief more than once. The affiliation flow I was handed had CNP inviting each club; on the design briefing we learned that clubs sign themselves up, so the flow started again from the other end.

Fourteen workflows, one file

The brief described fourteen workflows, from signing in to fines and reports. Each one got its own page in Figma: a lane per role, every screen named by flow, step and state, like W6.3 [default], and two frames that turn the page into a spec. The legend says how to read the page. The spec lists numbered acceptance criteria, so an issue or a test can point at a criterion instead of a screenshot.

The canvas became the source of truth for everything downstream. The flow diagrams the developers read were generated from it, never drawn by hand. When a weekly call changed a decision, the change went onto the canvas, not into a comment that would be lost a week later.

A system before the screens

Before the first flow, I set up the design system: Obra’s shadcn/ui kit, re-themed with CNP’s brand tokens and set in Geist. The developers build in React and Tailwind, so the components in Figma and the components in code share the same names and the same parts.

What the kit didn’t have, I built and moved into the library: a sidebar for each of the four roles, the stepper that runs through every wizard, document upload cards with their states, status badges, and the athlete row that appears on almost every screen. Every screen in the product is made of those pieces, which is what later let an agent edit the file without breaking it.

Four steps to a club

Affiliation is the first thing a club does on the platform, so it had to be the easiest. The wizard asks for a contact, then the club, then its documents, then a confirmation. The email is verified with a code instead of a link, which works on a phone, where the code fills itself in. A “same as contact” checkbox fills in the legal representative, so nobody types the same phone number three times. One line at the top says every field is required, instead of an asterisk on each one. Documents are PDF only, one card each, with their state on the card.

I tried the wizard with two club administrators before it went to the developers, and watched where they slowed down.

An entry, a fee, a gate

Entering athletes into a competition is where the rules pile up. The eligibility checks come first, before anyone picks an event, so a club finds out an athlete can’t compete before it fills in the rest. Classification is free text that CNP reviews, because every sport classes its athletes differently.

Every payment in the product, from the affiliation fee to a fine, goes through the same checkout: by card, or by bank transfer that a CNP operator checks within three days. One template meant one place to get payments right.

Who signs the referat

The referat is how a head coach asks for the budget to take a team to a competition. On paper it went from desk to desk, and each desk could send it back. In the platform it moves through four approval stages, the last of them optional, and the CNP operator corrects the request in place instead of returning it for a fix. The coach then signs it off with a checkbox and a timestamp.

A head coach walked me through a past referat, every stamp and signature on it, and the platform’s version follows the same path.

The agents in the room

Design was part-time on this project, and the agents are why that was enough. Issues for the developers were generated from the call transcripts, labelled and assigned as they were made, then checked by people in a weekly triage. Every time I changed a design, an agent posted a summary of what had changed on the issue, so a developer never had to compare screens to find it.

Before the enrollment page became a spec, I had it audited. Sixty-four agents read the page from different angles, checked each other’s findings, and returned 54 confirmed problems in twenty minutes: a happy path that ran through an error state, sample data that changed clubs halfway through, a CNP screen showing buttons only a club should have. The verdict was blunt: not a spec yet. The findings went back into the page.

Then an agent signed in as each role in a headless browser, walked the enrollment flow in the real app and found two backend bugs within minutes. Later, a role-by-role QA of the development app produced 39 screenshots, and an agent brought the Figma file back in step with the build, using only design-system components: nothing detached, every new colour and spacing bound to a token.

From the design briefing on 16 April to the first flows in development on 18 May took one month.

Where it stands

The platform is in testing with CNP and is being launched in phases, flow by flow.

The research system

Evidence about Kube comes from many places at once, and each one answers a different question:

• Support cases show what breaks for members and operators, in their own words. One study alone read more than thirteen thousand of them.
• Operator interviews and operator feedback show how operators run their locations and what they need from the product.
• User interviews show how members and prospects actually book, pay and get help.
• User testing checks a design with real people before it is built.
• Microsoft Clarity shows what people do on the portals, step by step through each flow.
• Stakeholder feedback brings in what sales, support and product colleagues hear from clients every day.
• QA walkthroughs on the test environment, app-store reviews, web analytics and competitor research fill in the rest.

Every piece of evidence is logged the same way: where it came from, a short quote, the theme it belongs to and how strong it is. A finding counts as strong when at least three independent records from two different kinds of source support it. A single support case, or one operator’s request, stays a signal until something else confirms it. Every count is a minimum and says what it was counted from.

AI agents do most of the gathering, reading and counting, under written rules: cite every claim, never fill a gap with a guess, ask when unsure, and publish nothing without my review. I set the questions, read the evidence and decide what it means for the product. Each initiative lives in a shared knowledge base that mirrors the team’s Confluence space, so designers, product managers and engineers work from the same evidence.

What it found in the member portal

The first big study put five sources side by side: the support cases, three QA passes I ran on the test environment, an operator’s feedback deck, internal feedback and my own notes. Fourteen themes came out of it, and a theme only made the list when more than one source reported it.

Two themes stood well above the rest, and one of them came up in every source. For confidentiality I’m not describing them here, but they set the plan: quick fixes first, then a larger redesign around the stronger of the two.

The study also corrected itself. When the agents rebuilt the counts, four of the five rates we had first published turned out to be wrong. The page kept the old numbers next to the new ones, with the reason for each change, so the team could trust what stayed.

Amenities that set a location apart

Operators list their amenities in their own words, so the same amenity appears under many different names. The agents cleaned up the amenity lists of 167 operators into 46 standard amenities and scored each one by how rare it is. A 28-source literature review, with every claim checked separately, covered how other products show rarity.

From that I wrote the ranking rule for a location page. Rare amenities, like a podcast room or a wellness room, lead, because they are what makes a location different. Amenities people expect, like Wi-Fi, parking and meeting rooms, always stay visible, because a location without them gets ruled out. Everything else follows by rarity. I designed the location-page mockup around that rule.

Finding a space on the map

Portfolio View lets a prospect browse all of an operator’s locations on a map before choosing one. I audited it, compared it with how other coworking platforms handle search across many locations, chose the discovery flow and drew wireframes that trace each decision back to a finding.

Before testing, an agent read the whole Figma prototype by script, more than 26,000 layers, and listed three missing interactions and the screens a participant should never reach. Our testing tool turned out not to handle prototypes that open new tabs, so we went ahead with one variant and moved the test to a working beta build.

Designing in the loop

The same setup also builds. For a partial payment feature, one agent wrote a brief for another: the goal, what to read first, what was already decided and the tasks. The second agent built the feature on the real checkout screen with the team’s design-system components, with three error states and recalculated totals. Halfway through I moved the action from inline editing into a modal. The build ended with six business rules the design could not settle on its own, written up for engineering.

Three jobs a day

Local employers, from Costa and Coop to Jersey Heritage and the Royal Yacht Hotel, post part-time and seasonal roles. Each person gets exactly three of them a day, and each one expires after 24 hours. The feed is open from 8am to 8pm.

The home screen is a small stack of three cards, one job each. The rate, the weekly hours and the days sit on the card, so it takes a glance to decide, and a bar under the stack counts down the time left to apply. Apply sends their profile straight to the employer, with no CV needed. View opens the full job: what you’ll do, what you need and where it is.

Matches before the sign-up

Onboarding starts with a few quick questions, from the kind of work someone wants to the parishes they can reach. Then three matches arrive one card at a time, before anyone is asked for an email. The sign-up comes after they’ve seen what they would get, with Google or Apple.

Lights out at eight

From 8pm to 8am the market is closed. The brief asked for that pause on purpose, and I gave it a screen of its own: a night sky over the island, a countdown to the next batch and a button to be reminded when it lands. Closing reads as a promise for the morning, with nothing left to scroll.

The employer side

Employers use the same app. The welcome screen splits the two paths, “I’m looking for work” or “I want to hire”, and a new employer account is checked within one business day before it can post. The dashboard leads with live jobs, new applicants and a button to post a job.

Employers review applicants, shortlist them and start the conversation. Job seekers can’t message an employer first, which keeps the inbox about the jobs they applied for.

One file, one system

The brief asked for forty screens, twenty-five for job seekers, twelve for employers and three shared, with loading, empty and error states. Everything lives in one Figma file, built from one component library: buttons, inputs, the job card, status pills, chat bubbles and the tab bar, which switches its tabs for the employer role. Every colour, radius and spacing is a variable.

The content is real Jersey: real employers, real parishes and real hourly rates, all in British English. Nudge is now in development.

The problem

Creators earning across YouTube, Twitch, TikTok, Patreon, and other platforms have no single place to understand their total income. Each platform has its own analytics dashboard with different metrics and different levels of API access. YouTube is the only major platform that exposes full revenue data through its API, while Twitch doesn’t share earnings at all, TikTok locks Creator Fund payouts inside the app, and Patreon only gives gross pledge amounts before fees. For creators treating content as a business, this fragmentation means missed tax obligations, poor forecasting, no way to identify which platform is most profitable, and no visibility into declining revenue until it’s already happened.

The solution

Kulma brings earnings, audience data and content performance from every major creator platform into one place. It uses a hybrid data model: automatic API sync where platforms allow it, and a streamlined manual entry system (quick add and CSV bulk import) for platforms that don’t expose revenue data. A trust indicator system labels every data point as verified, calculated, or manual so creators always know the confidence level of what they’re seeing. The design system was built on OKLCH color ramps with a violet primary palette, and every screen, component, and data visualization was designed for both dark and light modes.

The problem

Most recipe apps start from the wrong place. They show you thousands of dishes and leave you to figure out whether you actually have the ingredients. The result is a frustrating loop: find a recipe, check the fridge, realize you’re missing half the list, and either give up or make yet another grocery run. Meanwhile, perfectly good food sits in your kitchen until it expires because you couldn’t think of what to make with it in time.

The solution

Munchbase is a recipe platform built around how people actually decide what to cook. Every recipe comes with the details that matter upfront: preparation time, difficulty level, and dietary tags like high protein, vegetarian, gluten free, and more, so you can filter down to exactly what fits your lifestyle and schedule without scrolling through hundreds of irrelevant results. The homepage is a full screen vertical feed of recipes and chef content you can browse passively for inspiration, while a dedicated discovery section lets you search by cuisine, cooking time, difficulty, and dietary preferences when you have something specific in mind.

And when you truly don’t know where to start, Munch, the app’s built in AI kitchen companion, lets you scan your fridge, identify what you have, and generate personalized recipe matches ranked by ingredient fit, served as swipeable cards you can save or skip in seconds. A complete recipe database with smart filters, curated content, and an AI layer that meets you wherever you are in the cooking decision.

The problem

When I joined the project, I found a system that technically displayed data but gave researchers no way to change it. The existing dashboard showed tanks, corals, and nursery layouts, but every edit required downloading a CSV, opening it in Google Sheets or Excel, modifying the data manually, and uploading it back. Something as simple as moving one coral from one tank to another took 15 to 20 minutes. At a scale of over 20,000 coral fragments spread across 300+ tanks in three separate nurseries, each carrying complex metadata like species, genotype, growth stage, health status, and disease flags from AI powered camera monitoring, researchers were spending more time on data entry than actual restoration work.

Defining the structure

Researchers think systematically and spatially about the nursery: “Array A, Tank 3, Tray 19, slot B4”. The interface I designed follows the same structure, mirroring the building hierarchy so finding a coral on screen feels like finding it in the lab.

Each coral has a digital twin

Each coral in the system carries its own profile showing growth rate, mortality risk, and health status. Cameras mounted in each tank photograph the trays every hour, and an AI image recognition system analyzes the photos to update these metrics automatically. The researchers can see at a glance which corals are thriving and which ones are at risk.

Fixing a broken flow

I spent the first week on site purely observing: how the nurseries were structured, how corals moved between tanks and from ocean to land, what data researchers tracked, and where their workflows broke down. Most of the researchers kept surfacing one issue: moving a coral in the system took longer than moving it in the lab. So the work started with that one flow, and everything else followed. I mapped out the existing processes built around CSV files and manual updates, then designed new flows that eliminated the friction points I’d observed.

The new flow enabled the researchers to move corals in the system much faster. Select the coral, choose a destination, confirm. The system handles the rest, from logging the movement, updating the record, and maintaining the audit trail automatically.

Bringing the ocean into the interface

In-situ nurseries presented a different challenge. Unlike laboratory tanks, ocean sites are scattered across a coastline and researchers needed to see where things were geographically. The in-situ view combines a map with a list showing status and capacity for each site. Icons distinguish between structure types using the same visual language the team already uses in their planning documents.

Accessible art

Art galleries can be intimidating, without context contemporary art often feels hard to understand. Our solution was that each artwork in Artvue has a dedicated page with both audio and text descriptions. Visitors can play the audio while taking in the piece at their own pace, or read through the details if they prefer. The pages also include information about the artist, the style, the historical context, and connections to other works in the exhibition.

Scan to instantly identify any artwork

Wall placards are set in small type, mounted at awkward heights, and positioned so a few people reading them block the entire piece. In crowded galleries, you either crane over someone’s shoulder or skip the context entirely. Our solution was a scan functionality that lets users point their camera at any artwork from whatever distance. The system automatically detects which piece they’re looking at and takes them directly to the detail page.

An intuitive database

Sometimes visitors want to explore before they arrive, or find a specific piece they heard about. Our solution was a database search engine containing all artworks, artists, and collections in the gallery. Users can search by keyword or browse using filters like medium, style, period, or artist. The interface was designed to be simple and user-friendly, letting people quickly find what they’re looking for or discover something new.

An interactive map

Visitors often feel lost in galleries, unsure where to start or how to find specific pieces they came to see. Our solution was an interactive map showing the full gallery layout with every artwork marked. Users can tap on any piece to see its location, making it easy to find exactly what they’re looking for.

An AR-powered home for the arts

Most online art platforms treat artwork like any other product listing: a flat image, a title, dimensions, a price. Buying art asks for more. People need to feel something. They need to understand the artist’s intent, see how a piece would look in their space, and trust that what they’re buying is authentic. Kaleido’s marketplace was designed to feel like walking through a gallery. The interface stays minimal so the art can breathe, while giving collectors intuitive paths to discover artists across 150 countries.

Experiencing art in your space

With the Vision Pro experience, collectors can browse artists and artworks in a spatial environment, then place pieces virtually on their own walls to see how they fit before committing to a purchase. This closes the biggest gap in buying art online: not knowing how it will actually look and feel in your home. The interaction is simple by design. Point, place, resize, step back and look. There are no menus or settings to get in the way, so the decision stays with the art.

The full story behind every artwork

Collectors want the story behind a piece too. The desktop showcase screen was designed as a single, fluid scroll that gradually reveals everything a collector needs. It opens with the artwork front and center, then flows into detailed specs, additional shots from different angles, and a process timeline where the artist documents how the work was made, from first sketch to final piece. The scroll ends with a clear call to action to purchase. Every transition is smooth and intentional, giving the experience the pacing of a curator walking you through a private showing.

Three hours, four steps

The first half hour went on a quick look at existing book apps, to see how people search and browse and which features they use most. The second turned that into an app structure and a feature set. The third went on wireframes in Figma for the main flows. The last hour and a half went into a clickable prototype, with usability ahead of visual polish.