Avnzor
Three-phase engagement: a custom Shopify Plus theme rebuilt into a bilingual, accessible Hydrogen storefront — followed by a 181-PR post-launch audit, over which checkout completion rose to 27.5% and web orders 26.9% (observed, not yet split-tested).

Built and evolved Avnzor's commerce platform across three phases: first a custom Shopify Plus theme for code-free campaigns, then a bilingual English/Arabic Shopify Hydrogen rebuild audited to WCAG 2.1 AA before staging (34 issues cleared), and finally a post-launch silent-failure, speed, and admin self-service pass across 181 PRs. Over that 30-day window checkout completion rose from 25.7% to 27.5% and web orders 26.9% at a steady 127 SAR basket — observed alongside the work, not isolated by a controlled test.

The Constraint
Avnzor is a fast-moving health, beauty, and pharmacy retailer in Saudi Arabia doing 5,000+ orders a month across skincare, haircare, and supplements. In Phase 1, the problem was operational: the buying team depended on a developer for every campaign swap, every new landing surface, and every merchandising edit. Nothing on the storefront moved without code.
Phase 1 solved that bottleneck, and then the brand outgrew the theme’s architecture. There was no component model, no design token hierarchy, no Arabic-first layout logic, and no accessibility baseline. The roadmap ahead — a personalized supplement routine, bilingual search and recommendations, custom admin tooling, and Arabic treated as a primary layout direction rather than a translation patch — needed a clean rebuild on Shopify Hydrogen.
Phase 2 got the headless storefront live without taking a busy store offline. Phase 3 tackled what only shows up once real traffic hits a new stack: silent failures returning 200 OK while dropping writes, third-party scripts dragging mobile collection pages past five seconds, and merchandising surfaces that still needed an engineer to edit.

The System Across Three Phases
Phase 1 gave the buying team independence inside Shopify Plus. I built the theme as a kit of composable sections backed by structured metaobjects for ingredients, routines, and product relationships, wired into Klaviyo lifecycle flows. Editorial spacing and type rules lived in the sections themselves, so merchandisers could assemble campaigns fast without drifting off-brand.
Phase 2 moved the storefront onto Hydrogen and Vercel’s edge network. Three tracks ran together: semantic foundations (ARIA, keyboard focus, RTL parity), a three-tier token hierarchy (primitive → semantic → component), and the integration layer (Algolia for bilingual search and Frequently Bought Together, Clerk for accounts). Pairing Archivo for Latin with Almarai for Arabic handled RTL typography across every route without per-page overrides.
Phase 3 treated the live storefront like an operations room. Over a 30-day window (181 pull requests, 7 → 88 test files, 807 passing tests), I audited every silent failure along the purchase path, stripped megabytes of third-party JavaScript off initial page loads, and expanded the internal admin console from 53 to 79 routes (26 new self-serve screens with 16 owner guides) so marketing owns the Journal, announcement ribbon, gift guides, and deals rails directly.
The guiding rule across all three phases: the storefront is infrastructure, not a skin. If a form fails silently or a campaign needs a deploy, the job isn't finished.
Phase 3: Clearing Silent Failures After Launch
The biggest gains in Phase 3 didn't come from visual redesigns — they came from places where the storefront looked healthy in the browser while quietly dropping customer intent underneath.
Across 6,750 out-of-stock product pages, the Notify me when available form posted a customer email and a product tag to /newsletter. That route read the email, ignored every other field, and returned a green confirmation state — leaving 0 variant-tagged restock requests in Shopify. I rebuilt back-in-stock against variant-level storage with single-send restock alerts so demand on sold-out SKUs actually converts when inventory lands.
In the cart drawer, entering a held promo code like KSA96 (23% off with no minimum spend) hit a hardcoded fallback that told the shopper the code needed a bigger basket — sending people away to add items for a threshold that didn't exist. I moved discount rejection reasons server-side per code and added a rollback guard so trying a rejected code never wipes a working one already on the cart.
Three more quiet leaks showed up under inspection. Meta’s Conversions API had been switched on for months, running its payload builder on every page and POSTing to two gateway hosts that connect-src blocked every single time — invisible because standard pixel beacons still fired. On every product page carrying a specs last updated date, formatting without an explicit timezone rendered one calendar day on the server and another in the browser, triggering a React hydration tear-down in both locales. And on the homepage, a 23-slot experiment configured as an A/A baseline actually had one extra rail in Variant B, shifting every downstream slot index by one; I fixed the slot pairing and three self-linking Journal CTAs; over the same window hero click-through moved from 1.1% to 8.1%, though without a clean test I can't say how much of that the fix carried.
On speed and uptime, I replaced the 1.2 MB Intercom bundle (+251 KB avatar and a ~1 s phone input script loading on every route) with a zero-JS server-rendered launcher that loads Intercom only when tapped. Self-hosting Archivo and Almarai plus giving the homepage category quad a matching Suspense skeleton dropped homepage CLS from 0.36 to 0. Caching paid-social collection URLs (?utm_*, fbclid, gclid) at the CDN cut mobile collection TTFB from 2.86 s to 0.96 s and LCP from 5.45 s to 4.0 s. After an 18-minute Facebook link-preview flood pushed the 1 GB Micro database into swap, I put a 3-second timeout and stale-cache fallback in front of every storefront read so the site stays up even when the database stalls.
Key Decisions
Phase 1: Theme as a kit of merchandising primitives
Context
Avnzor’s buying team was small and non-technical. Every campaign change required a developer, making seasonal launches slow and fragile.
What I chose
Section-based Shopify Plus theme backed by structured metaobjects for ingredients, routine tags, and product relationships — decoupled from page templates.
Tradeoff
Higher upfront build effort because every section had to work as a reusable primitive rather than a one-off layout.
Phase 2: Custom Hydrogen app over stretching Liquid
Context
Personalized routines, bilingual Algolia discovery, and custom admin workflows pushed past what Liquid templates could hold cleanly.
What I chose
Full Shopify Hydrogen rebuild on React Router v7 with a three-layer token architecture (`primitive → semantic → component`) and WCAG 2.1 AA audited per route before staging (34 issues cleared).
Tradeoff
No native Shopify theme editor out of the box — solved in Phase 3 by building 26 seeded self-serve admin screens tailored to Avnzor’s exact rails.
Algolia with bilingual indexing over Shopify native search
Context
Shopify’s built-in search is English-centric and lacks a bilingual recommendation layer across thousands of Arabic-named health and beauty SKUs.
What I chose
Algolia search plus Recommend API for Frequently Bought Together, backed by sync pipelines that patch Arabic titles and metafields.
Tradeoff
Ongoing Algolia cost and dedicated sync scripts to keep Arabic index fields clean.
My Stack: a replenishment schedule, deliberately without a health score
Context
Customers buy four supplement bottles, take them unevenly, and forget to reorder. Category leaders like SuppCo drive engagement by scoring your stack out of 100 and flagging nutrient gaps.
What I chose
Built the scheduling, email one-tap adherence logging, and Refill Radar run-out prediction — and deliberately left out stack scoring and dosage gap claims, which cross into medical advice for a KSA retailer.
Tradeoff
Gave up the most shareable mechanic in the supplement category until counsel signs off on regulatory framing.
Phase 3: Server-rendered zero-JS support launcher & CDN-cached ad landings
Context
Paid social shoppers land cold on mobile collection URLs carrying `fbclid` and `utm_*` query strings, which bypassed CDN cache and loaded 1.2 MB of chat scripts before they saw a single product.
What I chose
Normalized marketing query params at the edge so paid clicks hit warm CDN cache, and gated third-party chat behind a static button until clicked.
Tradeoff
Proactive chat popups can no longer auto-fire on page load — a deliberate win on mobile, where the floating bubble had been sitting directly over the checkout button anyway.


How the System Changed the Work
Phase 2 delivered a production Hydrogen storefront on Vercel’s edge network. Every route was audited to WCAG 2.1 AA before staging — 34 issues across four severity batches, all resolved first. Homepage scores Accessibility 100, SEO 100 in Lighthouse; PDP scores Accessibility 96, SEO 100, CLS 0. The design system ships with a full token layer and extracted component primitives (Button, Checkbox, Dialog), all ready for dark mode and theme variation without class-level changes. Search is powered by Algolia in both English and Arabic. The PDP surfaces six custom metafields. The Frequently Bought Together component uses the Algolia Recommend API with bilingual product titles. Arabic renders via Almarai with zero inline overrides.
My Stack shipped as the personalization layer: a five-tap quiz that leaves a new account with a populated routine in under sixty seconds, a schedule across five time slots, one-tap dose logging from an email link with no session required, streaks, and a Refill Radar that predicts run-out and turns it into the reorder. Every entry point — quiz, product page, cart, post-purchase prompt, and the guest stack merged on login — writes through a single path, so there is no second way to create a stack item. The rule set that decides which products answer which goal is data, not code, so merchandising tunes it without a deploy.
Measured Result
Phase 1: The buying team now ships campaigns and new product surfaces without engineering involvement. The redesigned product pages are live, and the system became the foundation for the brand’s broader digital surface; no causal conversion figure is claimed without a controlled baseline.
Phase 2 baseline, measured at launch:
- Orders/month running through the rebuilt storefront, migrated without taking the store down
- 5,000+
- PDP UI/UX fixes in a single audit pass
- 19
- Lighthouse · Homepage Accessibility / SEO
- 100 / 100
- Lighthouse · PDP Accessibility / SEO / CLS
- 96 / 100 / 0
- Shopify collections root-caused as unpublished to Hydrogen
- 202
Phase 3, the 30 days to 15 Sep against the 30 days before:
- Checkout completion across the 30-day Phase 3 window (25.7% → 27.5%)
- +1.7 pts
- Web orders (3,719 → 4,720) & web net sales (+26.4%) at steady 127 SAR AOV
- +26.9%
- Mobile collection TTFB (2.86 s → 0.96 s) & homepage CLS (0.36 → 0)
- 0.96 s · 0
- Sold-out PDPs moved from a dropped /newsletter post to real variant restock alerts
- 6,750
- WCAG 2.1 AA issues resolved pre-staging · Automated tests passing (7 → 88 files)
- 34 · 807
- Self-serve admin screens running live (53 → 79 routes, 16 owner guides)
- 79
Why checkout completion leads instead of raw session conversion: Raw storefront conversion rose from 0.46% to 0.57% over the same 30-day window (17 Aug – 15 Sep vs 18 Jul – 16 Aug), which looks flattering until you inspect the daily series. 28 July in the prior window carried 81,971 non-shopping sessions against a ~25,000 daily norm while checkout entries stayed flat (543) — inflating roughly a third of the raw session-conversion gain (+0.07 pts when stripped vs +0.10 pts raw). Checkout completion (25.7% → 27.5%, 4,534 completed checkouts vs 3,575) isn't touched by top-of-funnel bot or campaign bursts, but it is still observational arithmetic, not a causal result. Two more limits: GA4 could not be read for this window, so every conversion figure rests on Shopify alone, and the window overlaps the previous monthly brief by 23 of its 30 days. The next conversion claim should come from a clean A/B run — the only homepage test so far was set up as an A/A and wasn't one.


What I'd Do Differently
I would run the Arabic catalog audit — metafield and title coverage across SKUs — in parallel with the Phase 2 build rather than after routing went live. Clean data before launch is always cheaper than backfilling against live pages. And I would enforce serverless bundle-allowlist tests from commit one: the day after the Phase 3 window closed (16 Sep), a dynamic import shipped without adding its file to vercel.json's includeFiles, returning 500 across routes alongside an empty-cart regression until I reverted both and added the allowlist test that now blocks that failure class in CI.
What's Next
Three clear levers sit open: upgrading the 1 GB Micro database instance so traffic spikes don't lean on the 3-second stale-cache fallback, running a clean homepage split test now that all 23 slots are properly paired, and working down the photography queue — 582 in-stock single-photo products that still sold 10+ units in 90 days despite thin imagery. On My Stack, turning computed Refill Radar run-out dates into one-tap subscription conversion is next while stack scoring waits on regulatory sign-off.




Let's talk about your project.
Avnzor — Product Taxonomy Rebuild