Empowering a Food Delivery App with AI: From Smart Tags to Meal Planning

Introduction

The food delivery space is growing increasingly competitive, and personalisation is now more critical than ever. Our team recently had the opportunity to design and implement a suite of AI-driven features for a food delivery client—each powered by OpenAI’s latest models, intelligent agents, and Python-based apps. These features were crafted to elevate the user experience through smart automation and semantic understanding of food items.

The goal wasn’t just to enhance the user experience—but to make ordering intuitive, insightful, and even fun. Below are the key innovations we brought to life:

1. Flavour Profiling: Quantifying Taste to Improve Recommendations
We developed a system to evaluate each dish based on its name, description, and image, quantifying it across five taste dimensions—Savory, Spicy, Sweet, Earthy, and Fresh. Using OpenAI GPT-4-turbo, the system returns a taste profile in JSON format:

{“Savory”: 85,”Spicy”: 10,”Sweet”: 5,”Earthy”: 30,”Fresh”: 70}

Example:

“Shrimp Scampi with Linguine” → Savory: 80, Spicy: 15, Sweet: 5, Earthy: 5, Fresh: 10

Uses:

• Powering flavour-based filters in the UI (e.g., “Show me fresh & spicy items”)
• Driving personalised recommendations
• Visualising items with radar charts or flavour wheels
• Ensuring robustness through error handling, even with missing data

2. Smart Tagging: Categorising Food with Precision

We implemented a tagging system using OpenAI prompts and business rules to auto-generate three types of tags:

• Visible Tags: Displayed to users (e.g., Poultry, Wraps)
• Hidden Tags: Internal use for logic and analytics (e.g., Entree, Side)
• Allergen/Diet Tags: Added when explicitly mentioned (e.g., Contains Eggs, Vegan)

Example:

“Sweet Shoyu Tofu” → Vegan, Contains Soy, Contains Sesame

Uses:

• Creating a clean, accurate taxonomy of menu items
• Preventing misleading assumptions
• Enhancing filtering, compliance, and personalisation

3. AI Review Generator: Turning Ratings into Words

We built a generator that creates short, natural-sounding reviews from ratings and feedback. It selects from over 15 “voices” (e.g., The Minimalist, The Food Enthusiast) and adjusts tone based on the rating.

Example:

Rating: 5 stars (Voice: Minimalist)
“Everything was spot on. Loved the grilled chicken bowl — fresh, filling, and flavorful.”

Uses:

• Transforming bland ratings into engaging content
• Enhancing authenticity and boosting UX
• Automating feedback processing at scale

4. AI Meal Recommender

Users can type what they’re craving, and the system returns matching dishes via the meal recommender agent.

Example:

Input:  “I’m craving something spicy.”
Output:

• Spicy Chicken Tikka Wrap – Bold flavours with a kick.
• Buffalo Cauliflower Bites – Crispy, zesty, and plant-based.

Uses:

• Ideal for indecisive users or those in a hurry
• Integrates well with homepages, voice assistants, or chatbots
• Can plug into health or mood-based personalisation engines

5. Conversational Order Assistant

We enabled a natural interface where users can ask questions like, “What did I eat last week?” The assistant uses past order history and the current menu to return intelligent responses.

Example:

User: “What did I order last Friday?”
Output: “You ordered the Grilled Chicken Wrap and Sweet Potato Fries.”

Uses:

• Human-like reordering experience
• Conversational UX for food delivery
• Aids retention by helping users rediscover favorites

6. Image-Based Food Search

Users upload a dish photo, and the image search agent identifies it using GPT-4o Vision, then recommends similar dishes.

Example:

Uploaded Image: Sushi Platter
Output:

Identified: Sushi Platter
Suggestions: Tuna Nigiri, Sashimi Combo, Dragon Roll

Uses:

• Enables discovery when dish names are unknown
• Powers camera-based ordering
• Enhances visual UX

7. Weekly Meal Planner

The meal planner agent generates a 7-day meal plan (breakfast, lunch, dinner) based on dietary preference and allergies.

Example:

Input: Diet – Vegan, Allergies – Dairy
Output (Monday):

Breakfast: Oatmeal with almond butter
Lunch: Quinoa and chickpea salad
Dinner: Lentil curry with brown rice

Uses:

• Simplifies health-focused meal planning
• Ideal for families and busy professionals
• Can drive subscription-based meal-kit offerings

8. Multilingual Ordering Assistant

Users input orders in English, and the assistant translates them into a selected language using Google Translate.

Example:

Input: “One large Margherita pizza, please.”
Output (French): “Une grande pizza Margherita, s’il vous plaît.”

Uses:

• Enhances accessibility in multilingual markets
• Simplifies localisation
• Improves customer support globally

9. Voice-Based Ordering

Simulates voice ordering using text as voice input. The assistant parses the request and updates the cart conversationally.

Example:

Input: “Add two chicken tacos and a Coke to my cart.”
Output: “You might be hungry! Adding chicken tacos and a Coke to your order.”

Uses:

• Powers hands-free interfaces
• Ideal for mobile-first or accessibility-centric apps
• Compatible with voice-to-text APIs

Behind the Scenes: How It Works

• OpenAI Agents: Each assistant is a custom GPT agent with carefully engineered prompts
• Async APIs: Ensures fast, non-blocking interactions
• Calorie API: Pulls nutritional data via CalorieNinjas or web scraping
• PDF Generator: Meal plans exportable via fpdf

Final Thoughts: Smart, Scalable, Delightful

These AI-powered features delivered measurable value:

• Faster decision-making
• Highly personalised ordering
• Multilingual and multimodal support
• Operational efficiency for the business
• By combining GPT agents, vision AI, and Python backends, we layered intelligence on top of a traditional food delivery app—delivering real, scalable outcomes.

Want to Build Something Like This?

If your team is exploring AI in food tech, personalisation, or customer experience, we’d love to connect. This architecture is flexible, future-proof, and ready to adapt to your platform.