Case study · 9 min read

Culinarr

Group ordering at UK food halls, where every person pays for exactly what they ordered.

Project
Product design projectSolo
Timeline
5 weeks
Tools
Figma, FigJamPrototyping
Role
Solo designerA 13-screen prototype across 4 flows, a nine-hall audit, four interviews, a five-participant usability test and a WCAG colour-contrast check

The brief

Culinarr is a prototype for group ordering at UK food halls. Scan a QR code, browse every vendor's menu in one place, and each person pays for exactly what they ordered. No app download, no account, no end-of-night bill chaos.

  • One menuEvery vendor in the hall, browsed in one place.
  • Your own shareEach person pays for exactly what they ordered.
  • Nothing to installNo app download and no account.

Problem

The problem

Anyone who's ordered food for a group at a food hall knows the drill: four separate queues, dishes arriving in waves twenty minutes apart, and a bill at the end that gets split evenly eight ways, even though two people didn't drink and one just had a salad.

I audited nine UK food halls to see what already existed. Four had QR ordering, and those covered one vendor or drinks only, never the whole hall. The other five were counter service with no QR at all. I couldn't find anything built for ordering across a whole hall as a group.

"A QR-triggered web app where one person opens a session, everyone joins on their own phone, browses across every vendor, and settles their own share. No app download, no account, no post-dinner payment chasing."

The product I set out to design

The behavioural research was more consistent than I expected. Splitting a bill equally tends to inflate spending by around 37%, 78% of diners say they'd rather pay only for what they ordered, and people are 3.2× more likely to overpay than risk looking awkward about money in front of a group. Asking for a fairer split costs more, socially, than overpaying does.

Sources: Gneezy, Haruvy & Yafe (2004); Splitty (2025) survey; my own nine-hall audit.

Existing tools get close but don't get there. In my comparison, me&u only partly supports spontaneous group sessions and filters dietary needs per vendor, and Uber Eats assumes one restaurant and one host who pays for everyone.

My starting hypothesis: if everyone could browse, order and pay for their own items from their own phone, with no app and no account, both the coordination friction and the payment friction would drop.

Research

Research

To get past my own assumptions, I combined behavioural-economics literature on group spending with a competitive audit of nine food halls and a read-through of reviews across TripAdvisor, Reddit, Google and Facebook. I also interviewed four people. Three findings shaped what followed.

Group dining behaviour research
01

Group dining behaviour

Sequential ordering sets an unspoken spending ceiling: whoever orders first anchors what feels acceptable for everyone after them.

Payment splitting psychology research
02

Payment psychology

72% of people find talking about money at the table harder than religion or politics. Fairness is wanted, but the social script for it just doesn't exist yet.

72 out of 100people find talking about money at the table harder than religion or politics.

What it came down to: a fair way to split a bill rarely gets suggested out loud, so the product needed to make that conversation unnecessary.

Direction

What the research asked for

How might we

help groups coordinate orders, pay fairly, and enjoy the meal without an app, an account, or one designated payer?

That changed the order of work: coordination first, payment second. Two decisions follow from it.

Users

Who it’s for

I used two personas, built from that research and the four interviews: Sam, a host who just wants everyone fed without becoming the evening's logistics manager, and Zara, a student with dietary restrictions who'd rather quietly filter a menu than ask a server to double-check every dish.

Structure

How it’s structured

Journey mapping, built from the research, is what reframed the project for me. I went in assuming payment was the painful part. The emotional low point was ordering: a host guessing what six people want over a WhatsApp thread, four have replied, two haven't, and everyone's getting hungrier.

Customer journey map
Customer journey map: the low point is ordering, not paying.
Information architecture
CulinarrOpened from a QR code at the table, in the browser, with no app download and no account
  • Join
    • Scan QR at the table
    • Enter 4-digit code
    • Open WhatsApp link
    • Enter your name (no account)
  • Session
    • Who’s at the table
    • Invite more (QR, code, link)
    • Shared cart, live
    • Host controls
  • Browse
    • All vendors
    • Vendor menu (Food · Drinks)
    • Search
    • Filter by diet
    • Quick filters (Quick Bites, Food in 10 mins)
  • Cart
    • Group cart, grouped by person
    • My items
    • Host submits the order
  • Pay
    • My share (itemised by default)
    • Equal split toggle
    • Payment method
    • Remaining table balance
  • Track
    • Status per vendor (Received · Preparing · Ready)
    • Group ready notification
  • Accessibility settings
    • Simplified menu view
    • Colour vision options
    • Dyslexia-friendly font
    • How order status looks

Solution

Shaping the solution

Two decisions shaped the solution.

Decision 01

Guests choose, the host submits

What
Anyone can join a session by QR, a 4-digit code, or a link shared on WhatsApp for people who arrive late, and add whatever they like. Everyone sees the shared cart as it grows, so nobody is guessing what the table has ordered. Only the host can send it.
Why, three reasons
  1. If anyone could submit, someone would send the order while others were still deciding, and the group would end up with several partial orders.
  2. The host is also the person the table already expects to coordinate, so the rule matches how groups behave.
  3. One submit moment locks the order, so every share is final before anyone pays.
Cost
The host holds the one lock on timing, so a distracted host can hold up the table.
Decision 02

Everyone pays for their own items

What
Itemised by default. Equal split is a toggle, not the assumption.
Why
Itemised won because the app already knows who ordered what, so the fair split needs no extra work from the group.
Cost
Shared plates aren't split among selected people. The hybrid option was too complex for the scope, which is what I gave up.

I compared four ways to split the bill

  • Host sets the splitGives one person too much power and latecomers no say.
  • Group voteAdds friction before anyone has seen the menu.
  • HybridPersonal items plus shared plates. Too complex for the scope, which is what I gave up.
  • ChosenItemised by defaultThe app already knows who ordered what.
Early wireframes
Early wireframes.

Design system

One system for every vendor

Colour
Primary actionsCulinarr redThe Add buttons, the vendor header strip and the live-session banner.
RatingsAmberThe rating badge on each vendor header.
Text and headersInkVendor names, prices and the dark photo gradient behind them.
ScreensSurfaceThe page behind menu cards and the search field.
DividersLineImage placeholders, outlines and quiet separators.
Type
InterHeadings, prices and buttons in the appGarlic Bread Bites £5.00
InterDish descriptions and helper textButtery, garlicky, dangerously addictive
Components
Primary button
Default · Hover
Filter chip
Quick BitesFilter by Diet
Quick filter · Dietary filter, labelled
Rating badge
4.3 ★
On every vendor header

Three directions, one kept

I tried a neobrutalist direction early on, with bold outlines and high contrast, and liked it for about a day. It made dietary tags harder to scan and read as trendy rather than trustworthy for something handling real payments. I dropped it for the calmer system above.

Early wireframesThe neobrutalist direction I droppedThe final prototype flows
One card system for every vendor

The hardest visual problem was making a burger menu, a sushi menu and a cocktail menu feel like one coherent product, not a generic third-party aggregator. I used one card component everywhere. Variety comes from each vendor's photography and colour header, and the structure never changes. The trade-off is that vendors get less control over their own layout.

Create and share a group session
01

Create

One tap after scanning generates a 4-digit code, a QR and a link, whichever gets the group in fastest.

Browse the unified menu
02

Browse

One menu across every vendor, with a cross-vendor dietary filter (text labels alongside icons) and a dyslexia-friendly font toggle.

Pay your own share
03

Pay

Everyone defaults to their own items. A live “remaining table balance” shows what's settled. One failed payment is designed not to block anyone else's order, though I haven't tested that against a real failure.

Track order status
04

Track

Each item moves through Received → Preparing → Ready, per vendor, and the group is notified when everything is ready. I didn't find per-vendor tracking in the apps and halls I reviewed.

Testing

What Testing Changed

Five participants, in person, SUS afterwards. Average task completion was 84% against my 80% target, ranging from 60% to 100%. Everyone completed creating or joining a session and paying. Four of five completed browsing and ordering, and tracking an order. Setting dietary filters fell short at 60%, so that's the one I'll walk through.
84%
Average task completion, against an 80% target
79.5
SUS (“Good”), against a target of 75
60%
Completed setting dietary filters
5
Participants, all university peers

Task completion by task

Create or join a session100%
Pay100%
Browse and order80%
Track an order80%
Set dietary filters60%

"I spent 40 seconds looking for the filter before I asked. Once I found it I used it immediately. It's just not where I expected it."

P3, during testing

The filter icon was being read as general settings, not dietary filtering. The fix wasn't a better icon. It was removing the need to guess at all.

Before
Before: icon-only dietary filter
After
After: labelled Filter by Diet button

Testing also exposed a gap in the split. P3 asked what happens if one person wants to pay individually while the other three split the rest, which my original design didn't cover. I changed it so the split mode can be switched even after someone has paid, and the remaining amount adjusts automatically.

SUS came out at 79.5 (“Good”) against a target of 75. With five participants I'd treat that as directional. The colour-contrast checks I ran all passed WCAG AA. The bigger limit is who was in the room: all five were university peers with high digital literacy. Older diners, people less comfortable with new interfaces, and anyone in a genuinely loud food hall weren't tested.

Prototype

Try it yourself

The final design, as a working prototype. Pick a flow to read what to try, then use the prototype below. The same file is linked at the end of this page.

Scan the QR on the table tent, enter your name and start or join a group session. No account, no download.

The Figma prototype opens in its own tab.

Open the working prototype (opens in new tab)

Validation

Where it landed

What the result shows
  • The core flows worked

    Everyone completed creating or joining a session and paying. Four of five completed browsing and ordering, and tracking an order.

  • Usability scored “Good”

    SUS came out at 79.5 against a target of 75, and average task completion was 84% against 80%.

  • The weak spot was fixed and retested

    The dietary filter fell short at 60%. After the labelled filter, confusion dropped by roughly 5%.

  • The limits are the people tested

    All five were university peers with high digital literacy. Older diners, people less comfortable with new interfaces, and anyone in a loud food hall weren’t tested.

What I’d do next

This was my first time running a structured usability study start to finish: recruiting, writing tasks, scoring SUS, coding interview notes. Five participants felt small the whole way through, until the same patterns kept showing up independently across sessions. If I ran it again I'd recruit six to eight before treating a finding like the dietary-filter failure as more than a strong hint.

  • An in-situ pilot

    Every test so far has happened in a quiet room with cooperative participants. A real food hall is loud, crowded, and nobody's full attention. That's the environment that will actually break things.

  • A bigger sample on scope

    The "Buy a Feature" exercise that shaped scope had three participants. Joining a session scored 0/3 even though I'd marked it a Must Have, so I made the join flow much more minimal. With three people I can't tell whether that was signal or noise. Six to eight would make it defensible.

If you'd like to see the full flow, here's the working prototype (opens in new tab), or feel free to reach out directly. Happy to walk through the trade-offs and the parts that didn't work.

open to opportunities

Let's talk design.

Open to roles in product design, service design, and UX, across consumer, hospitality, and complex multi-user products.