Operating model
What a 15% service fee actually leaves behind.
Every figure below is derived from one file of inputs. The inputs are published, split into what the operator specified and what is still only assumed, and each assumption carries a note saying what it rests on. If an assumption is wrong the arithmetic is still right — change the input and every number on this page moves with it.
One order
The goods pass straight through. Everything else has to be carried.
Nothing is marked up: the customer is charged what the item costs on the shelf in Korea, and that same amount goes back out when we buy it. So the goods net to nothing, and the service fee and what is charged for carriage are left to cover the domestic leg, the international leg, handling, a cross-border card rate, and the parcels that never turn up.
| per order placed | usd |
|---|---|
| Goods, at Korean retail | $57.60 |
| Service fee | $8.64 |
| Carriage, charged | $32.00 |
| …the goods, bought | −$57.60 |
| Korean domestic leg | −$2.50 |
| Carriage, paid | −$22.00 |
| Check, consolidate, pack | −$3.00 |
| Card, cross-border | −$4.62 |
| Parcels that do not arrive | −$2.39 |
| What is left | $6.13 |
An order is 3.2 items at $18.00, so $98.24 on the invoice once the fee and carriage are on it. 3% of parcels never arrive; the customer is made whole and the service carries it, so that sits against every order rather than the unlucky one.
Two ways to read it
An order leaves less than the fee the customer thinks they paid.
That is the whole shape of a buying service. The fee is what is being compared and negotiated, and it is not, on its own, enough — the margin on carriage is what closes the gap. Price carriage at cost and the service runs at a loss on every order.
The fee the customer pays
$8.64
15% of $57.60 of goods. The number a rival buying service is compared on, and the one a deck would lead with.
What the order leaves
$6.13
71% of that fee, and 6.2% of what the customer paid. The gap is closed by $10.00 of margin on carriage — price carriage at cost and this goes negative.
The library
The real cost centre is video, and it never stops.
This model is built on at least three shorts per product. The shelf today is not there — the products standing on it carry one each, and every figure below is what the catalogue this page describes would cost, not what has been spent. Those videos are not free and they do not stay useful — short-form creative decays, and a product whose best video is a year old is a product that stops being shown. Replacing a quarter of the library each quarter is not a marketing line item. It is the cost of goods for a business whose actual product is attention.
Cost of one usable videoA seeded creator keeps the product ($18.00 at Korean retail) and takes a fee. Not all of them post — this figure carries the ones who did not.
$130.00
What the current library cost to assemble1,847 shorts across 240 products.
$240,110
Videos replaced each monthA quarter of the library each quarter, so a product's best video stays roughly a season old.
154
Video refresh, per month
$20,009
Fixed overhead, per monthConsolidation space, platform, tooling. No salary.
$1,400
93%
of the monthly fixed base is video, not overhead. Every conversation about cutting costs in a business like this one starts with the space and the software, which together are the other 7%.
Break-even
Said per product, it is a number an operator can act on.
This is what each of the SKUs on the shortlist has to sell every month simply to pay for its own videos. A SKU below it is not underperforming — it is being subsidised by the ones above it, and the shortlist should be shorter.
Orders per month to break even
3,493
…which is gross revenue of
$343,137
Video cost carried by each SKU, per month
$83.37
Orders each SKU must produce, per monthBelow this, a product is being subsidised by the rest of the catalogue.
13.6
How it would be built
Medusa 2.0, with video as a first-class object.
Commerce is not the interesting part of this build and should not be written. What is interesting is that the thing being sold is a video, and a Medusa 2.0 module may not reach into another module's tables — so video cannot be a column on Product. It is its own module, associated through a link and read back across the boundary in a single query. The buying, consolidation and seeding flows are durable workflows because they are long, they call outside services, and a half-created order must never become visible.
Commerce modules, used as shipped
- Product
- The shortlist as a catalogue. A product exists because a curation entry was approved, not because someone typed it in.
- Pricing
- USD price sets. One currency: the service fee and the carriage are quoted in dollars wherever the customer is.
- Inventory
- Not used to hold stock. Variants are created with `manage_inventory: false` — we buy each unit after the order, so there is nothing to decrement and nothing to reserve.
- Stock Location
- One location, and it is not a warehouse: the Seoul consolidation point where goods bought on a customer's behalf wait for the rest of their order.
- Cart
- Built from a video, not from a grid. The line item carries the clip it came from.
- Order
- The order is the request to buy. What happens afterwards is tracked next door, in Purchase.
- Payment
- Cross-border card rates, which are higher than domestic and are not refunded on a return. The model treats them as sunk.
- Fulfillment
- One shipping option: everything leaves Korea in one box once the last item has been bought.
- Customer
- Accounts, addresses and order history.
- Promotion
- Launch pricing, drop codes, creator codes.
- Tax
- By destination country. In a buying service the customer is usually the importer, which changes who owes what — this is the module where that gets encoded, not a detail to settle later.
- Region
- One region, Global, covering the twelve countries we ship to today. Japan and Mexico are there because this business has actually sold into them.
- Sales Channel
- Web storefront now; the app feed and any TikTok Shop surface later, without a second catalogue.
- Auth
- Email and social sign-in.
- Notification
- Order mail, sourcing updates, the message that says an item was unavailable.
- File
- Creative assets. Video itself does not live here — it stays on the platform it was posted to.
Infrastructure modules
- Workflow Engine
- Redis. Durable execution: a long-running buying job survives a restart and rolls back on failure. In-process would lose it.
- Event Bus
- Redis. Decouples ingest and campaign flows from the request that started them.
- Cache
- Redis. Feed responses and price calculations.
- Locking
- Redis. Stops two ingest runs creating the same SKU twice, and stops two instances taking the same lock.
Data stores
- PostgreSQL 18
- Every commerce module and all three custom ones. Medusa's own store, unmodified.
- Redis 8
- Event bus, cache, locks, workflow state. Four modules, one instance, separable later.
- Object storage + CDN
- Product photography and creative. Not built yet.
- MongoDB
- Planned, not built: view and conversion time series, where the write pattern is append-heavy and the shape changes often. Nothing depends on it today.
Written for this product
Video
Clip
Every product is sold through short-form video and nothing else, so video is a first-class commerce object rather than an attachment. Linked to Product with `defineLink` and read back in one `query.graph` call.
Holds the embed reference, not the file and not a cached poster frame. The storefront embeds from the platform on viewport entry.
Curation
Source record, listing image, our own selection note
The shortlist is the product, so how something got onto it has to be written down. The note shown to a shopper is ours — no outside platform's ranking is presented as an endorsement.
The image credit field records where each photograph came from, so replacing it with the brand's own asset later is a data change.
Purchase
Buying job, buying line
A normal shop's order is finished when it ships. Ours has a middle — sourcing, purchase, inbound, consolidation — and each of those is a state a customer can reasonably ask about.
Lines carry their own status. Two of three items bought and one sold out is the ordinary case, not an exception.
Durable workflows
Shortlist an item
on: A candidate clears our own selection criteria
- 01Open a Curation record with the factual product data and the listing image
- 02Write the selection note — the reason a shopper will be shown, in our words
- 03Create the Product and its USD price set, with inventory management off
- 04Link Product to the Curation record
- 05Queue a seeding campaign for the minimum three clips
rollback Any step failing leaves the product in draft. A half-created SKU is never visible to a shopper.
Attach an approved clip
on: A creator deliverable is approved in SocialSeed.ing
- 01Record the deliverable against its seed
- 02Create the Clip from the platform's canonical URL
- 03Link it to its Product
- 04Recompute cost per usable video for the campaign
- 05Publish the SKU once it holds three clips
rollback A missing or deleted source video leaves the seed open and the SKU unpublished rather than shipping a product with a dead embed.
Run a buying job
on: An order is placed
- 01Open a Purchase job against the order, one line per item
- 02Buy each line at Korean retail and record what was actually paid
- 03Mark anything sold out as unavailable and refund that line, keeping the rest of the order
- 04Receive the goods domestically and consolidate them into one parcel
- 05Ship, and record charged carriage against actual carriage
rollback A failure part-way leaves the job in its last completed state with the bought lines recorded. Nothing is silently re-bought.
Inputs
Everything the arithmetic rests on.
Nothing else is typed anywhere. Assumed inputs are marked — they have not been validated against a real month of trading, and the ones most likely to be optimistic are the share of seeded creators who deliver usable video, the number of items in an average order, and a single carriage figure standing in for a table that really varies by destination.
Catalogue size
240
Operator's brief. Small on purpose: every SKU has to carry its own video cost, so the catalogue is a cost base rather than an asset.
Shorts in catalogue
1,847
Operator's brief. Above the three-per-SKU minimum, which is what a video-first storefront needs to avoid showing the same clip twice.
Minimum shorts per SKU
3
Operator's brief. Below three a product page repeats itself within one session, which is the point at which the format stops working.
Creators
120
Operator's brief. Campaigns are run through SocialSeed.ing, so the roster is a standing asset rather than a per-campaign spend.
Card rate, cross-border
4.4%
Standard online card pricing plus the cross-border and currency-conversion surcharges that apply when the buyer is not in the merchant's country — which, here, is every buyer.
Card fixed fee
$0.30
Standard per-transaction fee.
Average item, at Korean retailassumed
$18
What one item costs on the shelf in Korea, converted. This is what the customer is charged for the goods — the buying service does not mark the goods up, it charges for the service.
Items per orderassumed
3.2
Consolidation rewards larger baskets — one carriage charge covers the whole parcel — so buying-service orders run bigger than ordinary retail ones. Still a placeholder.
Service fee on goodsassumed
15%
What is charged for finding, buying, checking and consolidating. The visible price of the service and the number a customer will compare against a rival agent.
Carriage charged to the customerassumed
$32
A single blended figure standing in for a table that really varies by destination and weight. Set above cost deliberately; the model below shows why it has to be.
Carriage actually paidassumed
$22
Blended international rate for a consolidated parcel of this size. Japan is far cheaper than Mexico; a single figure hides a spread that pricing will eventually have to face.
Korean domestic legassumed
$2.50
Getting each purchase from the Korean seller to the consolidation point. Charged per order rather than per item because most items in one order arrive together.
Check, consolidate, packassumed
$3
Labour: opening, checking against the order, repacking into one parcel, photographing. The step a customer is actually paying the service fee for.
Parcels lost, damaged or refusedassumed
3.0%
Cross-border parcels go missing, get held, and occasionally get refused at the door. The service carries this, not the customer.
Seeding fee per creatorassumed
$60
Flat fee paid per seeded creator at the micro tier, on top of the product they keep.
Seeds that yield usable videoassumed
60%
Share of seeded creators who post something good enough to run as a product video. The number most likely to be optimistic.
Catalogue refreshed per quarterassumed
25%
Short-form creative decays. A quarter of the library replaced each quarter keeps a product's top video roughly a season old.
Fixed monthly overheadassumed
$1,400
Consolidation space, commerce platform, domain and tooling. Excludes any salary.
Rated Korean catalogueassumed
9,000
Order of magnitude for how many skincare products carry an ingredient rating at home. Used only to express the selection ratio; nothing else depends on it.
Raw inputs as JSON: /model.json · this page as Markdown: /model.md
Still shut
The shop opened around this document. Its own till did not.
This page is fully static, stores nothing, and has no account system — there is no database behind it to secure. Most of what this section once had to admit missing now exists: the backend is migrated and live, the catalogue is real — a shelf of products bought at Korean retail — and where a product also stands on Amazon's shelf, its page carries a door that works today. What is still not connected is this shop's own checkout: you cannot yet pay us. The arithmetic has been real throughout.