Building a Safety-First AI Cooking Assistant

In the world of AI-powered food applications, there’s no room for error when it comes to health and safety. A single mistake in allergen detection could have serious consequences. This is the story of how we built a sophisticated French cooking assistant that processes over 35,000 recipes with an unwavering commitment to safety, performance, and user experience.

The Challenge: Beyond Simple Recipe Search

When we set out to build our cooking assistant, we weren’t just trying to create another recipe aggregator. We wanted to solve the complex problem of personalized recipe discovery that could safely accommodate every dietary need, preference, and restriction. The challenge was multifaceted:

• Health-critical accuracy: Allergen detection must be 100% reliable
• Cultural complexity: French cuisine has intricate terminology and preparation methods
• Scale: Processing tens of thousands of recipes with consistent quality
• Real-time interaction: Users expect instant, conversational cooking guidance
• Cost efficiency: AI operations at scale can become prohibitively expensive

Our Dataset: The Foundation of Quality

Our system starts with a comprehensive dataset of 35,000+ French recipes sourced from Elle à table, one of France’s most prestigious culinary publications. This isn’t just a collection of ingredients lists – it’s a rich tapestry of:

• Complex ingredient expressions: “2 c. à soupe de farine” (2 tablespoons of flour)
• Culinary techniques: Detailed preparation methods and cooking instructions
• Nutritional information: When available, for enhanced filtering
• Seasonal data: Ingredient availability and optimal preparation times
• Cultural context: Traditional French dishes with regional variations

The diversity and quality of this dataset presented both an opportunity and a challenge. While it provided excellent raw material, the complexity of French culinary language meant that simple keyword matching would never be sufficient.

What We’re Building: The Complete Vision

Our cooking assistant is designed to be a comprehensive culinary companion that can:

1. Intelligent Recipe Discovery

Users can search using natural language, dietary restrictions, available ingredients, or even just a vague craving. The system understands context, intent, and constraints to deliver perfectly matched recipes.

2. Safety-First Personalization

This is where we differ from every other recipe app. Our system doesn’t just filter by tags – it performs deep allergen analysis on every recipe, considering:

• Direct allergen presence
• Hidden sources and derivatives
• Cross-contamination risks
• Processing method impacts

3. Adaptive Cooking Guidance

Once a recipe is selected, the system provides real-time cooking assistance through voice interaction, adapting instructions based on:

• User skill level
• Available equipment
• Household size
• Time constraints

4. Intelligent Substitution

When ingredients are unavailable or restrictions apply, the system suggests safe alternatives that maintain recipe integrity while respecting all dietary constraints.

The Architecture: A Multi-Layered Approach

To achieve this vision, we designed a sophisticated multi-layered architecture that balances performance, safety, and scalability:

Technology Stack: Modern, Efficient, Purpose-Built

Our technology choices reflect our commitment to performance, developer experience, and operational excellence:

Runtime & Foundation

• Bun: Ultra-fast JavaScript runtime replacing Node.js
• TypeScript: Strict typing across the entire codebase
• Monorepo: Organized with workspaces for maintainability

Backend & API

• Hono: Modern, fast web framework for our API layer
• Drizzle ORM: Type-safe database operations with PostgreSQL
• Zod: Runtime validation for API safety
• Qdrant: Vector database for semantic search capabilities

Frontend & User Experience

• React 19: Latest React with concurrent features
• TanStack Router: Type-safe routing with excellent DX
• TanStack Query: Server state management and caching
• Tailwind CSS: Utility-first styling with Radix UI components
• PWA: Offline capabilities and mobile app experience

AI & NLP Processing

• spaCy: Production-grade NLP with French language models
• Google Gemini: Multi-modal LLM for text and speech processing
• Custom NLP Pipeline: Hybrid rule-based and ML approach
• OpenTelemetry: Comprehensive observability for AI operations

Infrastructure & Deployment

• SST v3: Infrastructure as Code for multi-cloud deployment
• Docker: Containerized development and deployment
• Kubernetes: Production orchestration with auto-scaling
• AWS + GCP: Multi-cloud strategy for resilience

The Safety Philosophy: Deterministic Over Probabilistic

The most critical architectural decision we made was our safety-first approach to AI. While many systems rely entirely on LLMs for understanding and processing, we implemented a hybrid approach that combines:

1. Rule-Based Systems for Critical Functions

Allergen detection, dietary classification, and safety-critical parsing use deterministic rule-based systems. These are:

• 100% predictable and auditable
• Instantly verifiable for correctness
• Cost-free to operate at scale
• Culturally aware of French culinary specifics

2. LLMs for Enhancement and Validation

We use LLMs strategically for:

• Data enrichment and validation
• Natural language understanding in user queries
• Conversational interfaces and cooking guidance
• Complex reasoning about recipe adaptations

This approach ensures that critical safety functions never fail due to AI hallucinations or model drift, while still leveraging the power of modern AI for user experience.

Performance and Cost Considerations

Building a system that processes 35,000+ recipes and serves real-time user queries requires careful attention to performance and cost:

Processing Pipeline Efficiency

• Batch processing for LLM operations to minimize API calls
• Multi-level caching to avoid redundant computations
• Streaming architecture for real-time user interactions
• Background processing to keep the user-facing system responsive

Cost Optimization Strategies

• Smart caching at every layer of the stack
• Provider abstraction for competitive LLM pricing
• Usage analytics to identify optimization opportunities
• Efficient data structures to minimize computational overhead

The Development Experience: Docker-Native Workflow

We believe that great systems require great development workflows. Our Docker-native development environment provides:

# Start the entire stack with hot reload
make dev

# Connect to staging databases for real-world testing
make remote-dev

# Run comprehensive test suites
make test-all

This setup gives developers:

• Instant feedback with hot reload across all services
• Consistent environments from development to production
• Remote development capabilities with staging data
• Integrated testing across TypeScript and Python components