“There should be an API end point that allows authenticated users to send an excel listing of products and to have cart created and ready for them. Apparently a common use case for bars and restaurants.”
— The original request
Time to build a 100-item order
Before: one product at a time
58 min
After: upload, map, confirm
2–10 min
How I measured it: timed a 100-item order built by hand on the dashboard, then timed real files through the upload flow, from the cleanest (2 min) to the messiest (10 min).
Outcome220orders placed through the upload flow in its first month on the merchant dashboard. Storefront designed and specced, not yet released.
The order already exists, in a spreadsheet
Exported files are messy. An endpoint answers a messy file with an error array, which leaves someone holding a 167-line order and no idea which lines failed or what to do about them.
Key insight
I treated the upload as the easy half. The design problem was what happens to the rows that don't match, and whether someone can finish their order anyway.
Four steps, on purpose
The established pattern: file, map, validate, submit. Reusing it was deliberate. An import flow is the wrong place to be original.
Two entry points, one resolution path
Everything after the file is parsed is shared.
From inside an orderMerchant
Upload cart pageCustomer
Cart has items?Keep or remove
Map columnsBarcode, quantity, unit
Match catalogBy barcode
All matchedNothing to resolve
CartReady to check out
1
Some issuesMissing, multiple, OOS
Review and fixChoose or drop items, then the cart
2
None matchedDead end
Fix and retryExport failed barcodes, back to upload
Entry
Recovery
Result
1
Valid isn’t resolvedA barcode can match and still be ambiguous (Tito’s comes in seven sizes) or out of stock today. Neither is a data error, so a format check can’t catch it.
2
Recovery is a loop, not a repairYou can’t edit a spreadsheet into stock. Export the failed rows, fix them, upload again and choose Keep, so the rows already in the cart stay.
The same flow, twice
City Hive runs a strict design system that differs between buyers (Storefront) and merchants (Universal Dashboard). This design needed the same UX implemented on both platforms.
1
File: upload cart
Cart conflict is resolved before parsing. A merge you can't see is worse than a question you have to answer.
Dashboard
Storefront
2
Map: sort the file
Export headers are arbitrary (Item UPC, Order Qty, Pack), so mapping has to be manual. Knowing when you're finished doesn't. Submit stays disabled until the required Barcode field has a home.
Dashboard
Storefront
3
Validate: solve issues and approve
Three things can go wrong, and they're not the same thing.
Matches nothingA file problem. Export those rows and fix them.
Matches several sizesA choice. The buyer picks the right product.
Out of stockNot wrong at all. It's just unavailable today.
Dashboard
Storefront
Dashboard · nothing matched
Storefront · nothing matched
4
Submit: view updated cart
Back to where you started, with the cart filled. The merchant lands in the order they were already building; the customer lands in a cart ready to check out.
Dashboard
Storefront
Watch it run
Screen recordings of both flows, end to end: file, map, validate, submit.
Merchant dashboardShipped
StorefrontDesigned, not yet released
When the file doesn’t match
The two recovery paths: nothing in the file matched, and clearing an order to start again.
Nothing matched · dashboard
Nothing matched · storefront
Clearing the order
Merchant dashboard first
Storefront was specced and designed, not yet released. On the merchant dashboard, usage shows buyers adopted it, and two numbers show whether the recovery design held up.
First-try match rate41%
of uploads matched every row on the first try. The other 59% hit at least one problem row, so the review step carries most uploads.
Recovery loop74%
of rows that failed on an earlier upload now go through successfully. Failed rows aren't a dead end; recovery works.
Monthly usage220orders placed through the CSV flow in the first month measured, counted with a production database query.
What I'd take from this
Do differentlyTime the typical file, not just the extremesI timed the best and worst files, 2 and 10 minutes against a 58-minute manual baseline, but never a typical one. A median across real uploads would have turned the saving into one number instead of a range.
LearnedDesign the failure path firstOnly 41% of uploads matched on the first try. A flow built for the happy path would have failed most buyers. The review-and-retry loop is what made the feature work, and 74% of failed rows now go through.