Branch one
mayDisplayHalal(1) → false
- enNo halal wordingNone
- zh-HansNo halal wordingNone
- msNo halal wordingNone
Not certified, so no label
Client workWeb27 Jun 2026 to 19 Sep 2026
A three-branch restaurant's single-file site, rebuilt as a trilingual Next.js app. Three languages, one rule held by a test: an uncertified branch never shows a halal claim.
The greeting in three languages: Welcome. 欢迎光临. Selamat datang.
3 locales, 121 message keys each, no English fallback
Branch one
Branch two
Branch three
The halal rule
Halal claims are regulated in Malaysia, so a branch without certification must never show halal wording, in any language, anywhere. The UI asks mayDisplayHalal(id), and halal-rule.test.ts scans every locale clause by clause.
Certify a branch and its label appears in all three languages. Force the label onto an uncertified branch and watch the test.
mayDisplayHalal(1) → false
Not certified, so no label
mayDisplayHalal(2) → false
Not certified, so no label
mayDisplayHalal(3) → false
Not certified, so no label
Page weight
My first performance pass ran on localhost with no throttling and reported wins a phone never saw. So I profiled the deployed site on a throttled mobile profile, and fixed what it showed.
KB
Above-the-fold transferDown from 941 KB. Bytes through Netlify's image optimiser fell from 856 KB to 459 KB.
The reveal held content at opacity 0 until JavaScript ran. Driving it from CSS view() timelines fixed it, and 38 of 38 are visible with JavaScript off.
The navigation logo: 46 KB requested at w=1920 for a wordmark about 120 px wide.
21KB
Of render-blocking HTML, so they went.
Client work27 Jun 2026 to 19 Sep 2026
A three-branch restaurant's single-file site rebuilt as a trilingual Next.js app, with an owner-managed gallery and a test that stops an uncertified branch from ever showing a halal claim.
Built with an AI pair programmer, then reviewed and tested by me. The tests and decision records below are how I hold that work to account.
Rossi Italian has three branches in Malaysia: Penang, Kuala Lumpur and Ipoh. The first version of its site was a dependency-free HTML, CSS and JavaScript port of a design canvas, shipped with placeholder tiles instead of photos because the design-import tool truncated the originals. It could not carry three languages.
The rebuild needed real photography; English, Simplified Chinese and Bahasa Malaysia; booking per branch with no reservation backend; a gallery the client can change without a developer; zero recurring hosting cost; and one hard guarantee. Halal claims are regulated in Malaysia, so a branch without certification must never show halal wording, in any language, anywhere.
A Next.js 16 App Router site in React 19 and TypeScript 6, statically generated, with the gallery revalidated hourly. next-intl handles locale-prefixed routes and hreflang, and each of the 3 locales has 121 message keys. Booking is a Radix dialog that picks a branch and opens WhatsApp with a prefilled message: no backend, no stored customer data, and the channel diners already use. Motion uses native CSS scroll-driven animations gated on @supports and reduced motion, not scroll listeners or a new library. A Playwright loop screenshots pages for visual self-review, and every change is checked at 360, 390 and 768 px.
The application is 4,185 lines of non-test TypeScript. I built this with an AI pair programmer; every change went through tests and my review.
A git-backed CMS was the obvious way to let the client change photos. On this plan each save is a deploy costing 15 of 300 monthly credits, so about twenty photo edits could take the site offline. Cloudinary's shared credit pool was hard to forecast, and OneDrive share links return a viewer page. ImageKit won: one folder per branch, listed server-side and revalidated hourly, so a photo change costs one cached function call instead of a deploy, and image bytes skip Netlify's metered optimiser. Every failure path, from missing keys to a spent quota, returns a curated set committed in the repo.
A fallback that good hides integration bugs. ImageKit's listing is not recursive, so asking for the gallery root returns four folders and no photos, and the page would look the same because the fallback fills in. Each branch folder is queried directly, and failures go to the function log.
A documentation-only rule would regress the first time someone edited copy. A branch may show "Halal", "halal" or "清真" only when its halal field is certified. mayDisplayHalal(id) gates the UI, and halal-rule.test.ts scans every locale clause by clause and asserts that every certified branch actually says so. "Certification in progress" labels were rejected because they still imply a claim. When a branch was certified in September 2026, the rule needed no change, because it reads data rather than naming branches.
Eleven review passes found eight places where a governance doc stated something the code no longer did, and fixing them one sentence at a time never ended. docs.test.ts now fails the suite when the decision index and records disagree on status, when a doc cites a superseded record without saying so, when a doc states a halal status the data contradicts, or when a doc names an environment variable the code never reads. I mutation-tested it by restoring a stale line. It cannot catch a stale fact about code, and the menu proves it: the record saying menu categories link to the PDF is still marked accepted, although nothing links it.
A scripted edit of mine inserted an @media block by matching a focus-ring selector that appeared twice, so one copy landed between the two halves of a two-line selector. The dev server returned 500 on every locale, while tsc, ESLint, Vitest and next build all exited 0. I re-broke it on purpose to find out why: Turbopack drops an unparseable rule in a production build instead of failing. I had also been grepping build output for "error" rather than reading exit codes. Now css.test.ts parses the stylesheet with PostCSS and asserts that no selector has swallowed an at-rule and that braces balance, mutation-tested against the exact break.
The navigation's scroll-lock effect ran on mount with the menu closed, reset the body's overflow, and silently released the lock the intro cover had just taken. Each file was correct alone, and four per-file review passes missed it. I replaced both with one reference-counted useBodyScrollLock in a layout effect. The first integration test passed by accident. I found the discriminating case, closing the mobile menu while the cover is up, by reverting the fix and watching nine tests still pass.
My first performance pass ran on localhost with no throttling and reported wins a phone never saw. Profiling the deployed site on a throttled mobile profile, the largest image was the navigation logo: 46 KB requested at w=1920 for a wordmark about 120 px wide, because its Image declared a width of 1250 and next/image builds the srcset from that. Blur-up placeholders cost 21 KB of render-blocking HTML, so they went. The section reveal held content at opacity 0 until JavaScript ran: under throttling, 0 of 23 sections were visible at 500 ms. Driving it from CSS view() timelines made 23 of 23 visible, and 38 of 38 with JavaScript off.
101 tests across 17 files pass, measured with npx vitest run on 27 Sep 2026. data.test.ts pins every WhatsApp link to the branch data, after the contact section and the booking modal read numbers from two files and disagreed for one branch. A message check enforces key parity across locales.
The suite once printed "54 passed" and exited 1. jsdom implements neither scrollTo nor scrollBy, so a click handler threw an uncaught error instead of failing an assertion. Five of seven component suites also had no cleanup, and their assertions had adapted to leftover DOM with getAllByRole(...)[0]. A shared setup file made those assertions singular and passing, which proved the duplicates were the bug.
Reviews ran in lanes: React, TypeScript, security and WCAG 2.2 AA. They caught about 12 invisible menu controls in the keyboard tab order, now removed with inert; three small-text contrast failures, fixed with new --terra-ink and --terra-deep tokens beside the brand terracotta rather than darkening it; and a Next version inside a security advisory's range, upgraded from 16.2.9 to 16.3.5.
The site is live at rossi-italian.netlify.app, hosting costs nothing, and the client changes gallery photos without a deploy. Above-the-fold transfer dropped from 941 KB to 799 KB and bytes through Netlify's image optimiser from 856 KB to 459 KB. The project took 100 commits and 27 decision records.
What shipped, and when.
Rebuild from one static file to a trilingual app
A test stops uncertified branches showing a halal claim
Native scroll-driven animation, no JavaScript
Owner-managed photo gallery per branch
Above-the-fold weight cut from 941 KB to 799 KB
Live per-branch menus
7 decision records shown here, 1 of them superseded and kept on the list. Open one to read its context and consequence.
Decision record ADR-010 from 27 June 2026. Status: accepted.
Decision record ADR-017 from 28 June 2026. Status: accepted.
Decision record ADR-018 from 29 June 2026. Status: accepted.
Decision record ADR-019 from 11 September 2026. Status: accepted.
Decision record ADR-022 from 15 September 2026. Status: accepted.
Sampled 16 Nov 2026, 02:51 UTC. Failures are shown as they were read.
Rossi Italian site status
Not sampled yet
Source: rossi-italian.netlify.app
Paste a job description. You get a verdict, the evidence that matches it with citations, and what the sources do not show.
Open to senior AI and software engineering roles: on site or hybrid in Sydney, or remote for Australian and US teams.
Email meBook 15 minCVRossi Italian site status
200 in 4275 ms
Answered with HTTP 200, but slower than the 1500 ms budget.
Source: rossi-italian.netlify.app