ParcelProof by Jeejak

Indonesian online sellers lose Shopee and TikTok disputes because their packing photos turn out to be useless: too dark, the label cut off, the box not shown sealed. Our app JejakPaket records the photos; this prototype checks, right after packing, whether they would hold up.

Two checks run side by side. The basic check measures brightness, sharpness and the shipping barcode in Go, with no model and no cost. The Gemini check reads the whole photo set and reports what it sees in a fixed schema. In both, our server decides the verdict; the model never writes text a seller reads, and is never told which tracking number to expect.

Measured, not claimed

Gemini 3.5 Flash-Lite on 64 staged parcels (37 labelled weak by a person), prompt 2026-10-07.1, before any seller sees a verdict. The targets were set before this run.

100%precision: every "Bukti lemah" was right (35/35); target 90%
94.6%recall: weak parcels caught (35/37); target 90%
0%schema failures or "unsure" answers (0/64)
$0.0017per parcel at the Vertex AI list price; 2.1 s median
$0for the basic check: dark 4/4, blurry 4/4 on held-out parcels

The set is AI-generated, not real orders; real photos are measured next, in shadow mode. Full reports: docs/eval.

Try a sample parcel

Staged parcels with dummy waybills, made for our evaluation. No real buyer data.

Recent live Gemini checks

Every live call is stored in Firestore, so any Cloud Run instance can show it. Newest first.

Loading…

Try your own photos

Your photos get the basic check only. They are not stored and are never sent to any AI model, because a real waybill shows a buyer's name and address.