My Stack
A routine builder on the Avnzor storefront — five time slots, automatic placement, and a classifier deciding what even counts as routine-able.

Started at 35 stacks, 62 items, and only 10 logged days in week one — then grew to 220 stacks and 420 items once I stopped waiting on shoppers to open a browser tab every morning. That early gap between building a board and ticking it off reshaped the whole replenishment side: run-out dates now calculate off pack size across all 346 active items, and check-ins moved to one-tap links on WhatsApp.
Early Evidence — and Ninety Days Later
The first read after launch was 35 stacks built, 62 items in them, and only 10 logged days. That gap between stacks created and days ticked off was the real finding: people assemble a routine readily and forget to come back to a website just to check boxes. Any replenishment system that waited on daily web check-offs was going to sit idle.
If I were starting again I would make logging the first-class action rather than the second. The board is satisfying to build and easy to abandon; the daily tick is where retention actually lives, and on day one it sat one screen away from where people landed.
There was also a live gap in the classifier at launch. It had first run with narrower criteria, leaving stale exclusion tags on a chunk of the skincare range — and because tags win outright, those products could not be added until I re-ran the pass with a broader definition.
Ninety days in, listening to both findings changed the numbers: 220 stacks and 420 items scheduled, run-out dates calculated from pack count and slot frequency on all 346 active items without needing manual streaks, and reminders moved to a morning refill job plus one-tap check-in links on WhatsApp.
- Stacks built (35 in week one → now)
- 220
- Items scheduled across slots (62 → now)
- 420
- Active items with run-out dates
- 346 of 346
- WhatsApp cart recovery (shared rail)
- 17.0% · 6,799 SAR
- WhatsApp refill & ritual campaign CVR
- 7.0%
- Morning refill check
- 08:30 daily
Caveat on the WhatsApp numbers: the 17.0% cart recovery and 7.0% campaign conversion run on the same WhatsApp lifecycle rail as My Stack refill nudges, across all segmented buyers — not My Stack users in isolation.

The Constraint
A shopper buys a serum, a vitamin D and a cleanser in one order, then has to remember what to use when.
Avnzor sells supplements and skincare to customers who are, whether they use the word or not, building a routine. Nothing on the store acknowledged that. Products were sold as individual items, and the sequencing — morning or night, with food or before bed, together or apart — lived entirely in the customer’s head.
My Stack turns that into a board of five time slots. An item can sit in more than one, because a cleanser genuinely is a morning and an evening step. Workout is deliberately time-agnostic and sorts last — it is bound to training, not to the clock.
Built into the Hydrogen storefront: an add-to-stack action on cards and product pages, a schedule board, daily check-off, and replenishment prompts in cart.

The System I Introduced
The first real question was not how the board should look. It was which products are even allowed on it.
A serum belongs in a routine. A perfume does not. Neither does a lipstick, a hairbrush or a pack of cotton pads — they are bought from the same catalogue and look identical to a naive filter. Get this wrong in either direction and the feature reads as broken: too permissive and the board fills with things nobody schedules, too strict and half the skincare range cannot be added at all.
I resolved it as a cascade rather than a single rule. Explicit classification tags come first and win outright — set in batch by a Haiku pass over the catalogue, and editable by hand afterwards, so merchandisers get a free override without touching code. Below that sit a hard veto for the obvious non-routine categories, an ingestible signal backed by a supplement-vendor allowlist, and finally a topical routine-step signal.
The veto has a deliberate escape hatch. “Hair, Skin & Nails Gummies” reads cosmetic to any keyword rule, but it is a supplement — so an ingestible form overrides the veto. Most of the accuracy in this system comes from that one exception.
System Decisions
Place the product for them, then let them move it. Dropping every addition into Morning would have made the board a to-do list the shopper has to sort. Instead placement is inferred, cheapest reliable signal first: the product’s own usage directions are read before any keyword rule, because that is the label speaking rather than a guess — and it is parsed in both English and Arabic. Only if there are no directions do topical rules apply, then supplement keywords in a fixed priority order that resolves products matching two rules at once. A pre-workout is caffeinated and training-bound; workout wins.
Treat placement as convenience, not advice. Retinoids are placed at night because they photosensitise, and that is a real reason — but every suggestion stays editable. The system is scheduling a shelf, not dosing a patient, and the copy is careful about the difference.
Let people build a stack before they sign up. Asking for an account before a shopper has seen the value would have killed the feature. Guests build in session storage; on first login the whole thing is posted once as a merge and cleared. Nothing is lost at the account boundary.



The Constraint That Shaped the Architecture
Shopify OAuth and Supabase Auth are unrelated systems.
The storefront authenticates customers through Shopify. The stack lives in Supabase. There is no shared identity between them, which means the usual Supabase pattern — row-level security keyed to the signed-in user — simply does not work here. The database has no idea who the customer is.
So the tables carry row-level security with no policies at all, which locks out the public key completely, and every read and write goes through a server route holding the service-role key. Identity is the Shopify customer ID, resolved server-side only after the login check passes, and per-customer scoping lives in the queries rather than in the database's own rules. The browser has no path to Supabase whatsoever.
It is more code than the default pattern and it puts the burden on discipline in the query layer. Given two auth systems that will never know about each other, it was the honest option.
Let's talk about your project.
Shedd — Dubizzle / OLX Group