I Built an AI Chef That Understands Cultural Cuisine. Here's the Tech Stack.

Patrick Carney — Wed, 22 Apr 2026 12:50:39 +0000

How

I used GPT-4o, Replit, and a lot of prompt engineering to solve my own dinner crisis.

I'm a builder. When I encounter a problem, my first instinct is to code my way out of it.

The problem? Every night at 6pm, I'd stare blankly into my fridge, overwhelmed by the decision of what to cook. Existing meal planning apps felt generic—they didn't understand my cultural food preferences, my limited time, or the fact that sometimes I only have a microwave.

So I built Saffron: an AI Chef that creates personalized meal plans based on your real life, not a generic template.

Here's the tech stack that powers it.

The Core: Prompt Engineering is Everything

The secret sauce isn't fancy infrastructure. It's a carefully crafted system prompt that gives the AI a specific persona and strict output guidelines.

Saffron isn't just "an AI that suggests recipes." She's a "globally renowned AI Chef and empathetic culinary guide" with explicit rules:

· Prioritize authentic cultural ingredients over generic substitutions
· Adapt to the user's stated kitchen equipment
· Always include a cultural note about the dish's origin
· Suggest lower-carbon swaps for sustainability

This level of specificity transforms a generic LLM into a specialized tool that delivers consistent, high-quality results.

The Stack

· Frontend: React with Next.js, styled with Shadcn UI for a clean, warm interface
· Backend: Node.js API routes handling user inputs and OpenAI communication
· AI: GPT-4o with a structured output format (meal plan table + featured recipe)
· Database: Replit Database for user profiles and saved plans
· Hosting: Replit (built and deployed entirely within the platform)
· Monetization: Gumroad for simple, no-code sales

The Biggest Learning: Structure the Output Relentlessly

Early versions of Saffron gave me beautiful but unpredictable responses. Sometimes a full recipe, sometimes just a list of ideas.

The breakthrough was defining an exact output format in the prompt:

1. Warm Greeting
2. Daily Meal Plan Table (Breakfast, Lunch, Dinner, Snack)
3. Featured Recipe (Title, Cultural Note, Ingredients, Steps, Chef's Tip)
4. Conversational Wrap-up

This structure makes the response consistently usable and easy to parse on the frontend.

What's Next

I'm currently building:

· Visual food logging (snap a photo, get nutrition estimates)
· Voice-controlled hands-free cooking mode
· Pantry-first planning to reduce food waste

The Takeaway

You don't need a massive team or complex infrastructure to build a genuinely useful AI product. A clear problem, thoughtful prompt engineering, and a simple stack can get you 90% of the way there.

Saffron is live and available here: https://nutriplannerai.gumroad.com/l/NutriPlannerAi

Happy to answer questions about the build in the comments!

DEV Community: Patrick Carney

I Built an AI Chef That Understands Cultural Cuisine. Here's the Tech Stack.

How