Look: name the objects you see in English — running on a real iPhone

Translating menu text turned out to be the wrong feature — "anyone can do that" (Google Translate). The thing I actually want is vocabulary by looking: glance at a desk and see mug 머그컵, laptop 노트북, window 창문. That's the Phase 2 "물체 영어 라벨링" use case, and it's what makes this a learning tool rather than a lookup.

What landed

• Backend POST /api/label (commit abb0668): an image (base64) → Gemini 2.5 Flash vision → distinct objects, each with an English name and Korean meaning. gemini.ts got a completeVision() (image + text parts), plus parseObjects. Vision is Gemini-only for now (multimodal + free).
• iOS "Look" tab (commit c65d225): camera or photo → downscale to 1024px JPEG → /api/label → a list of {english, korean} cards. It's the new default tab — the others (Scan/translate, Coach/suggest) stay.
• The same camera/photo source from the Scan feature is reused; only the backend call differs.

Why it "just works"

Honestly, most of the magic is the model. Gemini understands the scene and names things — we trained nothing. Our part is the thin pipeline (capture → downscale → send) and a focused prompt ("name concrete objects, English + Korean, skip the background"). A few hundred lines + a good prompt = a real vision feature. That's the leverage of building on a frontier model.

Getting it onto the phone

The slog was network access, not the feature (full write-up in docs/troubleshooting.md): Bun.serve had to bind 0.0.0.0, the app had to stop using localhost, and the hardcoded LAN IP kept going stale as the Mac hopped networks (hotspot → WiFi). Fixed for good by addressing the Mac via its .local mDNS hostname instead of an IP.

Honest product note

The labeling tech is commodity — Google Lens, Apple Visual Look Up, and Meta's own on-glasses AI already do "what am I looking at." The phone version here is a validation prototype, not the product. The bet is the form factor (glasses = ambient, always-in-view, hands-free immersion rather than a hold-up-your-phone lookup) and the learning design (tracking words you've seen, spaced repetition, your weak spots). The demo being cool ≠ daily value — that's still the thing to validate.

Commits: backend abb0668f8d12596148656a966750a13bfc4b3c27, iOS c65d225fb006b74b28bc7103b5528e91b3255a51.