Inspiration

Digital recipes break when scaling serving sizes, and vague quantities like "a pinch of salt" leave cooks guessing. Thousands of great recipes are also trapped inside unstructured YouTube videos. We built The Worldwok Fork to combine global culinary heritage with AI precision.

What it does

Global Recipe Hub: Browse authentic global recipes across 6+ world cuisines.
Dual-Engine Scaler: Exact numbers scale in 0ms; vague amounts (e.g. "a splash of soy sauce") are dynamically re-estimated via AI.
AI Recipe Generator: Create authentic fusion dishes using few-shot grounded AI.
Video-to-Recipe Transcriber: Turn YouTube cooking videos into structured steps in <3 seconds.

How we built it

Tech Stack: Next.js 14+ (App Router), React, TypeScript, Tailwind CSS, PostgreSQL, Prisma ORM, Google Gemini API (gemini-1.5-flash), and youtube-transcript.
Scaling Math: $$\alpha = \frac{N_{\text{target}}}{N_{\text{base}}}, \quad q_{\text{scaled}} = \max\left(0.1, R(q_{\text{base}} \cdot \alpha)\right)$$
Grounding Vector Similarity: $$\text{Sim}(D_i, D_j) = \frac{\mathbf{v}_i \cdot \mathbf{v}_j}{|\mathbf{v}_i| |\mathbf{v}_j|}$$

Challenges we ran into

Scaling Subjective Quantities: Solved using client math for numbers and debounced async Gemini LLM calls for vague text.
Video Latency: Switched from heavy video frame analysis to YouTube timed-caption extraction, cutting latency by 95%.

Accomplishments that we're proud of

Instant Dual Scaling: Zero-latency number scaling paired with AI natural-language estimation.
Fast Video Transcriber: Ingestion and conversion of YouTube URLs to structured recipes in <3 seconds.

What we learned

Use deterministic code for math ($\alpha \cdot q$) and reserve LLMs for ambiguous natural language.
Rewriting video transcripts into imperative steps improves UX and ensures content transformation safety.