Two sides. One person. Obviously.

The failure, precisely

A card reads 陳志明 · 業務部經理 · 台灣積體電路製造股份有限公司 on the front, and Chih-Ming Chen · Sales Manager · TSMC on the back. That is one human being with one job at one company, printed twice for two audiences.

Scan it with almost any card app and you get two contacts. Or you get one contact and the other side is silently discarded. Or the app never asks for the back at all, and you find out months later that half your address book is missing its phone numbers, because the direct line was printed on the English side.

This is the single most-requested fix in this category and it is still unshipped. One app in Japan paywalls back-side scanning behind a ¥600 monthly subscription. Most simply do not model the concept of a card having two faces.

Why it is genuinely hard

• Two faces are two images with no shared text. There is often not one matching character between them, so naive text similarity finds nothing to join.
• Han unification. 田中 is the same code points in Traditional Chinese and Japanese. Nothing in the bytes tells you whether that is *Tanaka* or *Tián Zhōng* — you have to infer it from phone country codes, address script and whether romaji appears.
• The back is not always a translation. Sometimes it is a company profile, a QR code, a product list, or a completely different person at the same firm. Assuming a merge is as damaging as never merging.
• Phone numbers conflict rather than duplicate. The Chinese side has the switchboard, the English side has the direct line and an extension. Naively deduplicating loses one; naively keeping both produces four numbers where there are two.

How Kardhaus handles it

1. It asks for the back

   Immediately after the front capture, before you put the card down. Not buried in a menu — a prompt over the still-live camera, with "yes, shoot the back" as the default focus and "no" exactly as easy to press.

2. It detects the script on each face

   Traditional Han, Simplified Han, kana, hangul, Latin — per face, not per card. That classification is what makes the rest possible.

3. It decides what it is looking at

   Same person in two languages, two different people, or a person plus a company information page. If it genuinely cannot tell, it says so and asks you rather than guessing.

   This is the rule that matters: never produce two contacts from one card, and never merge two people into one.

4. It cross-fills instead of overwriting

   Chinese name and English name both persist, in `name` and `nameAlt`. Same for the company. You do not lose 台灣積體電路製造股份有限公司 because TSMC was shorter.

5. It classifies the phone numbers

   Work, mobile, fax and extension are separate types, reconciled across both faces — so a switchboard on the front and a direct line on the back become two correctly-typed numbers, not four guesses.

6. It shows you where every field came from

   Tap a field and the region of the card image it was read from highlights. You verify rather than trust. Low-confidence fields are flagged and sorted to the top, with the warning kept on the record after you save.

Common questions