All work

Product Design  ·  2026

Student Visa Express

A study permit app for students applying to Canada on their own, and the design system I built so it stays consistent, with Claude Code as my build partner.

Role
Design Lead, ImmigrateAI Global
Scope
Design system, product design, analytics
Timeline
Sep to Nov 2026
Status
Live

An immigration app built on a generic template, with no design system, used by stressed students who often read English as a second language.

Built the Foundations v4 design system and a live brand guide, redesigned the applicant journey end to end, and set up a way of working with Claude Code that keeps every change on-system.

One system across the whole product, every answer feeding a red flag or the statement, and the team's first honest read of where applicants stop.

99onboarding questions wired to a flag or the statement
37document types, each with its own stamp
9pull requests from one student's test session
1designer, working in the real codebase with Claude Code

An officer's read, before the officer

Student Visa Express helps international students build their own Canadian study permit application. It asks about their situation, reads their documents, flags what an immigration officer is likely to question, drafts the statement of purpose, and puts everything into one package they submit themselves.

The people using it are under real pressure. A school start date, family money and a future in another country all ride on one file. Many of them read English as a second language. So every screen gets read slowly, under stress, by someone who cannot afford to misunderstand it.

A product that had outgrown its look

When I joined, the app worked but did not hang together. It had been built quickly by several people and an AI website builder, each adding screens their own way. Colours, buttons and wording drifted from page to page, and every new screen made it worse.

  • About 99 onboarding questions, and only 22% of applicants finished them
  • An audit showed most answers never reached the statement of purpose or a red flag, so people were typing things the product then ignored
  • Meaning was carried almost entirely by text, for an audience that often reads English with effort
  • No shared rules, so each AI-assisted change could undo the last one

One screen, three versions

The dashboard as I found it, after my first pass, and now. The first pass brought the brand in but kept the old habits. The system is what made it calm.

Dashboard before I joined: a grey admin template with a sidebar and five progress rings Dashboard, my first pass: gold next step, blue progress bars and four kinds of card Dashboard now: next steps on gold and a stamp for each stage

One next step, and a stamp for each stage of the file.

The dashboard, rebuilt from the app's own components with sample data. Hover the cards.

apply.immigrateai.global/dashboard
DashboardRed Flag Scanner® 10Document Manager 8Document BuilderPackage AO

Welcome back, Amara. Here is your study permit file.

Your study permit file

Every step below is read the way an IRCC officer reads it, so you see the gaps before you submit.

Next steps

Finish your answers

3 of 58 answers still to give. Your documents and your red flags both come from them.

Continue
  • Upload your academic transcriptsYou are studying now, so your transcripts show the officer your current programme.
  • Upload your letter of acceptanceEvery study permit needs one. We read the school and dates off it for you.
16 more after theseSee everything outstanding
Your progressStage 3 of 6
Your answers3 answers still to give
  • Documents · 8 still to upload
  • Red flags · 10 to address
  • Statement of purpose · Not started

Nothing is submitted to IRCC from here. You submit on the IRCC portal.

Red flags foundView all
622
  • Nothing in your file shows what you are returning toCritical
  • Your funds have not been checked against the current requirementCritical
  • Money that is not yours needs a letterCritical
IRCC processing time

6 weeks

From Canada · sample figure

Statement of purpose

Not started

Start drafting

Designing with Claude Code

I used Claude Code as my main tool: not to generate screens for me, but to build what I designed in the real codebase, check it against live data, and show me the result before anything shipped. The design calls stayed mine. The AI did the legwork, and I reviewed every change before it was committed.

  1. 01LookQuery the live database and the real screens before proposing anything.
  2. 02ProposeTwo or three options as working pages, each with a recommendation.
  3. 03DecideI choose, often with a correction. "Too heavy." "Three and a half steppers."
  4. 04Build and reviewReal code, tests and checks. I review it before it is committed.

The part that made this work was writing my decisions down. Every rule I settled went into three plain files in the repo: one for the product and its users, one for the design tokens, one for the rules. Every AI session reads them first, so a decision made on a Tuesday still holds the following month, whoever or whatever is writing the code.

  1. AuditLive data, the questions, ten reference videos
  2. Foundations v4Tokens, rules, the live brand guide
  3. Lawyer portalThe reviewer side, on the same system
  4. UploadsMany files at once, one loading visual
  5. Funnel analyticsThree pull requests, live the same day
  6. Demo dayOne student’s session, nine fixes

Foundations v4

The design system came first, because nothing else would hold without it. I took the app back to a small, strict palette that feels like paperwork rather than software: cream and white grounds, ink text, gold for the one main action on a screen, rust for warnings, and purple kept for one job only, the legal register.

Few choices

Two neutrals and nothing between them. One gold button per screen. Five corner sizes. Fewer options meant fewer ways for a page to drift.

Measured

Every pairing is checked for contrast. Card edges were tuned until they measured 3:1 against the cream ground, not against the white card, where they looked fine and were not.

Live, not a PDF

The brand guide is a page inside the app, built from its own stylesheet and components. If the guide looks right, the product does too.

Ground#F6F1E6Every page
Ink#131200Text and progress
Gold#FFD950One main action, the stamp
Rust#C7692ACritical findings
Dark#2E2820The nav bar
Purple#7B5EA7Legal only
Light purple#D9C9ECMissing, and legal review

The signature is one stamp. Gold, a perforated edge, tilted a few degrees. It marks something finished, a stage, a document on file, a group of answers, and nothing else, so it reads without English.

From first answer to a finished file

The product follows the order an applicant actually works in, and each stage ends in a stamp. The dashboard and the scanner are shown above.

Most applicants open SVE on their phone first. Every screen is built for 375px before it is built for a desk.

The dashboard on a phone
The Red Flag Scanner on a phone
Smart Questions on a phone

“Answer once”

Six groups of Smart Questions under one stepper. Every answer feeds the documents, the red flags and the statement, so nothing is typed twice.

apply.immigrateai.global
Smart Questions with the six-group stepper
The same step on a phone

Following the numbers

The team asked where people were dropping off, and nobody could say. So I mapped the journey from live data, proposed the funnel steps for review, then built the analytics into the admin portal with tested database changes and event tracking. It shipped as three pull requests in a day.

The first reading was honest and useful. Fewer than half finish onboarding, and most who stop do so at the same group of questions, "Health and record". That gives the next round of design a precise target instead of a hunch.

What was hard, what landed

Challenges

  • Two codebases with unrelated histories and 155 conflicts between them. Work had to be replayed into the company repo by hand until we moved everything to pull requests.
  • An AI website builder whose instructions were capped at 10,000 characters. It kept forgetting rules, and once replaced the brand plane with one it generated.
  • A dashboard that said "File review: Done" on a file nobody had read. On an immigration application, false reassurance is the one thing the product must never give.
  • Answers that went nowhere: 9 of the 14 statement topics the questions produced had no section to land in.
  • Payments run live, so checkout could never be tested with a real card.

Wins

  • A live brand guide that reads its own rules, so designers, engineers and AI work from one source.
  • Every answer wired to a red flag or a statement section, including a new section for admissibility.
  • "Done" now means a scan really ran. Not started, done and a count of open findings read as three different states.
  • Real IRCC processing times, synced daily, replaced a placeholder figure.
  • One student's test session on a demo day became nine focused pull requests the same day.
  • Funnel analytics shipped in three pull requests, giving the team its first honest read of where people stop.

What I learned

Working this way, one designer can carry an idea from a sketch to a tested change in production. The speed is real, but it only pays off with judgement and written rules behind it.

PRODUCT.mdDESIGN.mdAGENTS.md
Write it where the AI reads it

A rule that lives only in a conversation is gone by the next session.

// No celebration and no congratulation.// Reward is earned and specific.
Delete old rules

A stale comment kept steering the code until I removed it.

My notes about itThe page itself
Make it look at the real thing

Built from my notes, the first mock missed most of what mattered.

Measure before redesigning

The funnel pointed at one group of questions, not the whole flow.

Next project
Multi-Product Onboarding