Cooking Inspiration Made Easy with Llama 90B Multi-Modal

1. Explore Step: Engaging with Stakeholders & Defining the Problem

Deciding what to cook can sometimes feel like a never-ending puzzle—especially if you open the fridge only to see random ingredients staring back at you. After chatting with friends, family members, and fellow cooking enthusiasts, we noticed a recurring theme: people often lack inspiration or get bored with the same old recipes.

Can Generative AI Add Value?
We thought, “What if AI could snap a picture of your fridge’s contents and suggest meal ideas on the spot?” This would:

  • Save Time: No more endless debates on what to cook.
  • Add Variety: Introduce fun new dishes that users might not think of themselves.
  • Boost Creativity: Offer fresh, unexpected menu ideas for those “just not feeling it” days.

Our goal was simply to explore whether we could make a lighthearted, curiosity-driven app to spark creativity in the kitchen.

2. Design Step: Prototyping the Solution & Gradio UI

Why Llama 3.2 90B (Multi-Modal)?

We used Llama 3.2 90B, a multi-modal model capable of handling both text and images. This model’s large parameter count and image-recognition capabilities were key for identifying fridge items and proposing recipes.

Building the Gradio Interface

We kept things super simple:

  1. Upload a Photo: Users snap a pic of the fridge or pantry items and upload it.
  2. Automatic Recognition: The AI analyzes the image to detect common ingredients.
  3. Menu Suggestions: In seconds, users see suggested meal ideas tailored to the recognized items.

By leveraging Gradio, we got a clean, user-friendly interface up and running quickly, allowing testers to upload photos, see real-time results, and share immediate feedback.

3. Implement Step: Integrating User Feedback

We invited a small group of home cooks and hobby chefs to try the prototype over the course of a few days. Here’s what we learned:

  • Food Recognition Accuracy: The AI correctly identified around 80% of the ingredients. Mistakes popped up here and there—like mixing up lemons for kiwis or salmon for carrots when it was lying sideways in a bag.
  • Relevance of Menu Suggestions: About 65% of testers said the model’s recipes were in line with their real cooking habits. They found it genuinely helpful.
  • Fun Factor & Inspiration: The remaining 35% liked the creativity aspect but said they’d only rely on it fully when they had no cooking inspiration or wanted something playful.

We refined the prototype by adding clearer ingredient labels and a simple interface prompt asking users to confirm or edit any misidentified items before generating their meals.

4. Evaluate Step: Measuring Product Impact

Time & Creativity Boost

Testers felt that having quick, AI-generated ideas freed them from the common “what to cook” conundrum. With the app handling the brainstorming, they could spend more time actually cooking—or just relax and enjoy the process.

User Adoption & Next Steps

  • Adoption Drivers: Users were more likely to turn to the app when stuck for ideas or seeking a dash of fun in their routine.
  • Room for Improvement: Strengthening the model’s accuracy to handle more nuanced or obscured ingredients would make it even more reliable.
  • Potential Features: Future enhancements could include dietary preference filters (vegan, gluten-free, etc.) and a built-in review system for suggested recipes.

Conclusion

Our “satisfy curiosity” project proved that a multi-modal AI model like Llama 3.2 90B can provide an entertaining and practical solution to the everyday dilemma of “What’s for dinner?” By recognizing fridge items with reasonable accuracy and serving up creative menu ideas, this lighthearted app encourages users to explore fresh and unexpected meals—no culinary rut required.

Back to blog