Kuharica

Recipes on social media are hard to cook from: the amounts are in a voice-over, a caption or a comment, often in cups and in another language. Kuharica takes the link (or a photo of a cookbook page) and writes a real recipe: ingredients in grams and millilitres, clear steps, times and oven temperature, in English, Croatian, Bosnian or Montenegrin.

How it works

• The pipeline (Python): downloads the post with yt-dlp, transcribes the speech with Whisper, pulls stills from the video and reads the caption, the creator's comments and any recipe on a linked page.
• The writing: a language model turns all of that into a recipe that follows a strict JSON schema and a set of rules: never invent amounts silently, mark estimates, and say plainly when a post has no recipe in it. A validator checks every recipe and sends mistakes back to be fixed.
• The app (Next.js, React, Drizzle and MySQL): cooking mode with big text, automatic timers and the screen kept on, a weekly meal plan, a shopping list that adds the same ingredient up across recipes, servings that rescale, "what can I cook?" from what's in the fridge, folders, ratings, notes and friends' cookbooks.
• On the phone: installable as an app, in the iPhone share sheet through a Shortcut, with a push notification when new recipes are ready.
• Accounts and credits: free sign-up with daily credits; credits are only spent on recipes that work out.

It runs as a Docker Compose stack (app, database and the recipe worker) on my own server.