よいちかぞく 夜市家族

We started by asking Japanese travelers who had visited Taiwan what they actually came to night markets for. Twenty-two respondents, two short surveys: one on familiarity, one on stall preference. Two numbers came back clean.

From the survey groups we drew one persona. Not because one person stands for twenty-two, but because the team needed a shared face to design against. Yamada is who we sketched on the wall.

Below, the four challenges acted out as a single evening. Six beats, one traveler. The journey map in comic form.

For most signs, generic OCR holds up fine. Point the camera, extract the characters, hand them to a translator. The characters arrive. The food doesn’t.

OCR reads these four characters cleanly. Google Translate gives back “Donshan Duck Head.” A Japanese tourist still has no idea what this stall is: a self-serve braised-stew counter from Tainan’s Donshan district, where you point at items in a warming tray and they’re tossed in spiced soy broth to order.

That gap is what the case is about. Yamada doesn’t need a better OCR pipeline. He needs the cultural context the name is shorthand for: what form of dish it is, how it’s served, what commonly goes in it, whether anything in it might break his dietary rules.

So we built that context. We walked through three night markets, collected forty-eight photos across thirty-two food categories, and annotated each one by hand: name, region of origin, cooking method, common ingredients, dietary flags. The model doesn’t just recognise the characters. It pulls the cultural entry behind them.

Before we built, we scanned the field. Two categories existed already: generic translation apps and Taiwan night-market travel apps. Each had what the other lacked.

We mapped each feature on two axes: how directly it solves Yamada’s pain point (horizontal), and how much emotional surprise it brings (vertical). Two patterns came out clean.

The lower-right quadrant (Essential Foundations) is crowded. Every translation app and every night-market travel app already covers it. We had to build the baseline, but it wouldn’t set us apart.

The upper-right quadrant was empty. That’s the wedge: features that meet the same pain point but bring emotional surprise. Domain-specific image recognition tuned for night- market signage. Cultural context attached to every dish, not just a literal translation. That’s where 夜市家族 placed itself. The dashed arrow below traces the iteration path: build the baseline, then push the wedge.

We wanted the food-intro screen to feel like a single confident voice, not a scrapbook of borrowed images. A shared visual language across every dish mattered more to us than photo realism. So we built the image library ourselves.

Thirty-eight studio-style AI photos across savoury, sweet, drinks, and a fourth category for the ingredients we flag as common dietary triggers. One lighting style, one camera angle, one plating style. The image Yamada sees in the app holds together dish to dish, even when the stall version varies.

One wordmark, one palette. Both running through every screen of the app.

Three lessons came back from the four-month cycle. One about method, one about scale, one about how a designer and engineer team actually ships something together.

A class collaboration between NCKU’s Industrial Design and Computer Science master’s programmes. Three members on each side of the handoff.