Problem

Home cooks rely on recipes found across blogs, PDFs and screenshots, but these are often messy, inconsistent, and hard to adapt. SousChef, a cooking app that guides users step by step with visuals, voice and timers, wanted to move into a more AI‑driven phase. To do that, they needed a way to convert any ordinary recipe into a structured, SousChef‑ready format that could be reused across their ecosystem for search, personalization and future AI features.

Research  

I analyzed recipes from a range of websites to understand how ingredients, instructions and metadata appear “in the wild”. I looked for missing times, vague or ambiguous wording, inconsistent units and formatting. This led me to frame the challenge as a data‑transformation problem: first define a clear, strict output schema for SousChef, then use LLMs plus validation and repair steps to reliably fit any messy recipe into that structure.

Solution

I designed an AI Recipe Transformer as a three‑step workflow inside Streamlit: add, transform, export. Users can upload files, paste text, or fetch recipes from URLs. The LLM then condenses and parses the recipe, validating it against SousChef’s output specifications and repairing issues where needed. Finally, users preview and download the structured outputs as JSON or CSV, ready for analysis tools and other SousChef features. Throughout the flow, I kept the interface simple enough for non‑technical home cooks, while enforcing a consistent data format in the background.