KollectionCloset

What isn't real: label OCR. A hosted page can't run it — and testing with a web OCR engine would understate what Apple Vision does on the same photo, so it'd be a misleading result. The camera captures a genuine photo; the text it "reads" comes from a sample tag or from typing. Everything downstream of that text is the real algorithm. The shipping QR is a stand-in too — a live one is issued by the carrier against his account. The shipment number beside it is real, and really is unique.

9:41 ▮▮▮ ◗ ▰

KollectionCloset

Your closet is
worth something.

Check a brand in seconds. Build a box. We’ll tell you what we’ll take before you send it.

0 brands on the buy list

· ·

Fill the frame with the brand tag

What does the tag say?

This is where the OCR result would land. Type it, or pick a sample — including deliberately messy ones, to see how the matcher copes.

What the label says

Reading the label…

Detecting text

Identifying brand

Checking our list

Nothing in the box yet.

Sign in, or start an account

Same thing either way. Put in your email and we’ll send you a link. There is no password — nothing to remember, and nothing of yours to leak.

We never ask for a password or a card number.

Check your email

We sent a link to you. Open it on this phone and you’re in. It lasts an hour.

Nothing there after a minute? Check spam. New senders end up there until enough people say otherwise.

Create your account

So we can send your label, keep track of your boxes, and pay you. One screen, and you only do it once.

Name Email

—

How would you like to be paid?

Anything we can’t take comes back to you free. You can choose to donate individual pieces instead when you pack your box — it is a decision per item, not one setting for everything you ever send.

We never ask for a password or a card number. You can change any of this later.

—

Your shipments

Ready to send

We’ll email your QR code. Show it at the counter — no printer needed, and the postage is on us.

We never ask for a password or a card number.

Your shipment number

KC-000000

Sample code — the live one comes from the carrier.

The brand list says whether we buy it. This says what condition it has to be in. Worth two minutes before you pack — anything that fails here gets sent back or recycled, and that helps nobody.

Leave these out

StainsAny mark at all, including faint yellowing at the underarm. A stain that survived one wash will survive ours.

Holes, tears and snagsIncluding moth holes you can only see against the light, and pulled threads on knitwear.

Broken or missing hardwareMissing buttons, a zip that sticks or splits, a snapped drawstring, a broken clasp.

Smoke or mildewIt does not wash out and it transfers to everything around it in the box.

Heavy pillingBobbling across a whole panel — typically the seat, the underarms or the inner thigh.

CounterfeitsIf you are not certain it is genuine, leave it out.

These are fine

Creasing from packingComes out on a hanger.

Fine

A loose threadIf you can trim it, trim it.

Fine

A cut-out care labelCut for comfort is fine. The brand tag itself needs to be there.

Fine

Light sole wear on shoesTread still clearly defined, no heel collapse.

Fine

Washed but unwornIdeal, actually.

Fine

Judgement call

Light pillingA few bobbles under the arms. On a premium coat, fine. On a basics tee, not worth the postage.

Depends

Slight fadingBlack cotton going grey is common and forgiven on workwear, not on eveningwear.

Depends

Altered or hemmedTrousers taken up are fine if the hem is clean and even. A home job with visible stitching is not.

Depends

A mark that might wash outIf you are not sure, wash it first and look again. Do not gamble a slot on it.

Depends

How to check in ten seconds

Hold it up to a window — thin patches and moth holes show.
Turn the underarms and collar out. That is where things die first.
Run a hand over the seat and inner thigh for bobbling.
Smell it. If you have to think about it, it smells.
Work every zip and button once.

Not sure

We couldn't read a brand we recognise from that photo.

Tag cut out, worn away, or stamped into the leather? Tell us what it is.

Brand

What is it?

Search brands

Check your brands. Photograph the tag at the collar or waistband, or type the name. We tell you straight away whether we buy that brand and what we pay for it — before you pack a thing.
Get a free label. Build a box, then tell us where to send the money: name, email, and Zelle, Cash App or PayPal. No password. We email you a shipping label — post it or drop it off, postage is on us, and there is no minimum.
We check it and pay you. Once your box lands we go through it piece by piece and send what we owe you. Anything we can’t take comes back free, or goes to a local charity — you choose that per item when you pack the box.

Being on the list isn't a guarantee. Condition, season and how much of it we already hold all matter. The list tells you whether the brand is worth bringing in.

Your privacy

Photos are read on your phone and never leave it, and we keep no record of what you look up. Checking a brand is anonymous whether you have an account or not.

An account is only needed to send a box, because the label, the parcel and the payment all have to hang off one record. It asks for your name, an email and where to send your money — nothing else. We don’t pass them on, and we never ask for a password or a card number.

Your box

The box you are building is kept on this phone until you ask for a label. Clear your browser data and it goes with it.

Getting paid

You see the price before you send — it is on the brand, not haggled afterwards. Once your box lands we confirm each item against what you told us, and pay by Zelle, Cash App or PayPal, whichever you picked when you got your label.

Anything we can’t take either comes back to you at no cost, or goes to a local charity on your behalf — whichever you picked for that piece when you packed the box. It is per item, so a coat can come home while a worn-out tee goes to the charity shop.

What donating means

You choose per item, so you can have a coat posted back and let a worn-out tee go. Whatever you mark donate goes to a local charity, and the name and date come to you with your payment.

We don’t sell it and we don’t keep it as stock. If something you marked donate turns out to be worth buying, we ask you first and pay you for it. Passing on an item to keep it would be a good way to lose a seller for good, and we’d rather just make you an offer.

Still to confirm

The warehouse address — coming soon. It will be on your label, so you won’t need to write it down.

The brand list

What we buy shifts with what sells. Worth checking again before a big drop-off — a brand that was a no last month might not be.

For the demo

The owner's side, so you can see both halves without a second account. It would not sit here in the real app — it would be a separate sign-in that customers never see.

KollectionCloset · version 1.0 (preview)

Demo only. No sign-in, no security — anyone with this link can open it. Changes are saved in this browser and nobody else sees them. The real version is a separate login against the database.

The list

0 brands

0 accepted

0 rejected

Placeholder data — a plausible buy list for a mid-market reseller, and the prices are guesses too. Both come from the source rather than this browser, so his real list replaces them in one place.

How the matcher works

This is the part that ports straight to Swift or Kotlin unchanged, and the part worth getting right.

Text is normalised — lowercased, punctuation stripped.
Every brand carries aliases: each way the name is actually printed on a tag.
A second pass collapses the characters OCR habitually confuses — 0/o, 1/l/i, 5/s, rn/m — so 1u1u1emon still matches.
Then edit distance on individual words, for the rest.
Strong hit answers outright. A weak one becomes Not sure with suggestions, never a confident wrong answer.

Typing, with typos

Search is a separate pass — a few characters someone is still typing, not a blob of OCR text. It tolerates the mistakes people actually make:

Swapped letters. raplh lauren, vuroi, shien — the most common typing error there is, and the reason this uses Damerau rather than plain edit distance.
Wrong or missing letters. adiddas, nrth face, carhart, colombia.
Initials. tnf finds The North Face.
The other country's spelling. under armor finds Under Armour.

A clean prefix always outranks a fuzzy hit, so four letters into lulu you get lululemon — never a near-miss on something else. When a match came through an alias, the row says which one, so it isn't a mystery.

Where the admin went

There was an on-device admin panel here — a passcode, a hidden gesture, an Admin tab — and it has been taken out entirely. The brand list is now a constant in the source, and the app never writes to it.

Because the honest version needs a server anyway. A passcode that ships inside the app is visible in view-source and readable in a binary’s strings; it hides a panel, it does not secure one. That was tolerable while the list lived on one phone and the worst case was somebody editing their own copy. It stops being tolerable the moment the same screens reach other people’s email addresses, payout handles and money owed.

So his half is a separate sign-in against a real database, on its own address, with none of its code in the app customers download. Same data, two front doors. The schema and the endpoints for it are in BACKEND.md.

What the server holds

Profiles, shipments and the buy list all move to a database — that is what the Account tab is for, and it is the reason a label can exist at all: a carrier key cannot live in a browser, so something has to sit between the app and Shippo.

The box itself deliberately stays on the phone. A half-packed box is a draft, not a record, and syncing one would mean writing to a server on every tap for something nobody else will ever read.

Four tables and six endpoints, in supabase/schema.sql and netlify/functions. The app runs without any of it today.

Accounts, and what they are for

An account is required before a box can be sent, and it earns that: the label, the shipment and the payout all have to hang off one record. Without it every parcel is anonymous and the only way to answer “where is my money” is to search an inbox.

No password anywhere. Sign-in is a link by email. A password is a thing that can leak, and somebody selling a coat twice a year will not remember one — so the app asks for an email, a name and where to send money, and nothing else. It never asks for a card number, and it should not start.

Scanning still forgets. Nothing records what anybody looked up, whether they have an account or not.

Logo matching, realistically

Reading text and recognising a swoosh are different technologies. Text recognition is a free, instant, on-device API. Logo recognition is one of four things, and they are not equivalent:

Cloud vision API — ~$1.50 per 1,000 scans, needs a network, and its brand coverage skews to global giants. It knows Nike. It won't know Vuori or Alo, which are where the margin is.
Multimodal LLM — accurate and flexible, but a few cents a scan and a second or two of latency. Constrain it to pick from the brand list or say nothing, or it will confidently invent a brand.
A trained model — free per scan, offline, and needs 50–200 photos per brand. That's the wall most of these projects hit.
Image feature-print match — the one I'd build. iOS has VNGenerateImageFeaturePrintRequest: it fingerprints an image and measures distance between fingerprints, on-device and free. Needs three to five reference photos per brand, not hundreds. A new brand means uploading photos, not retraining.

All four need reference photos he doesn't have, and with nothing leaving the phone the app can't gather them for him. So they get added by hand, three to five shots of each brand's real tags, from whoever maintains the list. Collecting those is the long pole on any of the four, which is worth knowing before picking one.

What I'd check next

Tags with no text. A woven logo and a size, nothing to read. That's image classification, a different and much harder problem — and it's why the Not Sure path needs to be good rather than an apology.

Whether scanning beats typing. Three letters into the search box already answers the question. The camera has to be faster than that, or it's the demo and search is the product.

Category comes from the brand record, not the photo — nobody should accidentally build garment recognition.

KollectionCloset — working prototype

Your closet isworth something.

What does the tag say?

Reading the label…

Sign in, or start an account

Check your email

Create your account

Your shipments

Ready to send

Leave these out

These are fine

Judgement call

How to check in ten seconds

Not sure

Your privacy

Your box

Getting paid

What donating means

Still to confirm

The brand list

For the demo

Your closet is
worth something.