Cartaly

Photograph a restaurant menu you can't read and get back one you can order from — translated dish names and an AI-generated picture of every dish. Built for the moment you're sitting in a restaurant abroad and the menu is a wall of foreign text.

Live at cartaly.app.

How it works

photo → sha256(bytes) → cache hit? → dishes JSON (instant)
                            │ miss
                            ▼
              gpt-5-mini (vision, structured output)
              parse + translate in one call (~2–3 s)
                            │
                            ▼
              cards render (text first, image skeletons)
                            │ per dish, lazy, near viewport
                            ▼
              image cache hit? → R2 URL, else generate → R2

Design rules: text renders before images, nothing expensive is computed twice (deterministic cache keys everywhere), images generate lazily, and failures stay per-dish — one bad image never breaks a menu.

Features

• Menu scan — snap or upload a photo; the readable menu appears in seconds
• Dish photos — an AI-generated impression of every dish, loading as you scroll
• Search — diacritic-insensitive, across both languages ("pho" finds "Phở")
• Order builder + show-to-waiter — tap dishes into a list, then show the original-language names in big type
• Table share — every scanned menu gets a link; one person scans, the whole table browses

Full feature list in docs/features.md.

Stack

Next.js (App Router) + React on a single always-on Bun server, bun:sqlite via Drizzle ORM as source of truth and cache index, Cloudflare R2 for image bytes, OpenAI for parsing (gpt-5-mini vision) and image generation (swappable to Together via IMAGE_MODEL). No Redis.

Getting started

Requires Bun.

bun install
cp .env.example .env   # set OPENAI_API_KEY; R2 vars are optional in dev
bun run db:migrate
bun run dev

Open http://localhost:3000.

bun test        # run tests
bun run lint    # lint

Deploy

Docker Compose runs the app with a persistent volume for the SQLite database, behind an external tunnel-net network (e.g. a Cloudflare Tunnel):

docker compose up -d --build

Project layout

src/
  app/         routing shell only (pages, layouts, API routes)
  server/      server-only zone (AI calls, DB, actions, blob storage)
  components/  UI components
  lib/         browser-safe helpers
  schemas/     shared Zod schemas (client + server)
  config/      shared constants
docs/          spec, features, system design, implementation plan

All API endpoints return a standard response envelope — see src/schemas/standard-response.ts and src/server/create-response.ts.

Docs

• docs/SPEC.md — technical spec and pipeline
• docs/features.md — product features
• docs/system-design-overview.md — product overview
• docs/implementation-plan.md — build phases and current scope