LullabuyJudges start here

for the parent meeting a stranger in a parking lot

The money doesn't move until the camera has seen the item.

Buying a used crib or car seat? Your Visa payment is held, not sent. At pickup, your phone reads the label and scans the barcode against 1,208 real CPSC and NHTSA recalls and the banned product types. Clean: the seller gets paid. Recalled or banned: the hold is reversed and nobody takes the risk home.

Deal #0426 · pickupHELD

$64.00

Harppa high chair, model BHC001

MODEL NUMBER

BHC001

PRODUCTION BATCH

202408

Visa authorization held · checked at pickup

RECALL 26-061

Animated example: a payment is held, the label is scanned, then it is captured or reversed.

a Visa hold, not a promise

How it works

Visa's own auth-then-capture, the way hotels and gas pumps use it

  1. Step #1

    HELD

    Agree and hold

    You agree on a price. Visa authorizes the card and holds the money. The seller sees it is real; you have not paid.

    held, not paid

  2. Step #2

    SCANNING

    Meet and scan

    At pickup, point your phone at the label on the back or underside. The barcode scanner reads the UPC.

    we read the label, not the listing

  3. Step #3

    CHECKING

    Check it for real

    Model and batch go against every CPSC nursery recall; the photo goes through our trained banned-type model.

    no inclined sleepers, ever

  4. Step #4

    CAPTURED / REVERSED

    Capture or reverse

    No match: the hold is captured and the seller is paid. Recalled or banned: the hold is reversed. No dispute.

    and nobody took it home

Illustration (generated). The real scan runs on /pickup.

try it, it is the real index

Live recall check

Type what is printed on the label, or plug in a USB barcode scanner and scan the UPC. This calls the same/api/check the pickup scan uses.

Or try a real recall:

your verdict lands here

sees what the listing text hides

A model that knows a banned product on sight

Sellers rename things. A “baby nest” is an in-bed sleeper; “crib bedding set” can hide a bumper. We harvested 1,662 real marketplace listings and 243 CPSC recall photos, reviewed 496 labels one by one on numbered contact sheets, and trained a head on CLIP image embeddings. It is tested on 171 photos of products it has never seen: no recall and no listing sits on both sides.

2 of 107

ordinary baby items wrongly flagged by our model

38 of 107

wrongly flagged by off-the-shelf CLIP. A pre-screen that cries wolf gets ignored.

0.72

macro-F1 on held-out products (95% CI 0.60 to 0.81), vs 0.56 zero-shot

SCAN, ROUND 1 · 1,662 REAL LISTINGS

39 flags

We read every one: 1 real, 3 unsure, 35 false alarms (crib skirts, rail covers, dollhouse cribs).

learned from its
own mistakes →

ROUND 2 · SAME LISTINGS

9 flags

After adding 23 reviewed false alarms as training data. The held-out test set was not touched.

Per class, on unseen products

  • Inclined / in-bed sleepern=9

    precision 0.83 · recall 0.56

  • Crib bumpern=24

    precision 0.83 · recall 0.42

  • Drop-side cribn=31

    precision 0.95 · recall 0.68

  • Everything elsen=107

    precision 0.80 · recall 0.98

high precision, modest recall: it is a pre-screen. The label check at pickup decides.

Confusion matrix

Inclined / in-bed sleeperCrib bumperDrop-side cribEverything else
Inclined / in-bed sleeper5004
Crib bumper110112
Drop-side crib002110
Everything else020105

Rows: true class. Columns: model's answer. Numbers read live from /api/stats.