Halal Meals Rethinks AI: Focusing on Domain Expertise, Not Just Code

Zvonimir Fras’s work with Halal Meals underscores a crucial point in industrial tech adoption: the technical sophistication of the algorithm is rarely the limiting factor. Often, the true bottleneck resides in domain-specific data availability and operational scope.

The original narrative—that Halal Meals had an 'algorithm problem' requiring refinement—was quickly corrected by recognizing the real constraint: menu breadth and supply chain depth. Fras’s insight here is highly valuable for other SMEs looking to layer AI onto established, physical businesses. He notes that many companies treat advanced algorithms as a substitute for addressing foundational business issues ('AI as a placeholder for magic').

His methodology is fundamentally about 'extract[ing] pure AI from the things that don’t need to be AI, and use[ing] AI as the glue between those systems.' This shifts the focus of AI implementation from solving complex mathematical problems within the model, to intelligently connecting disparate operational data points (e.g., linking seasonal ingredient availability to dietary preference constraints). In a food service context, this means using AI not just for recommending 'taste' but for optimizing supply chain inputs and maximizing menu variety under real-world material limitations.

This approach—the concept of 'AI as the glue'—is pragmatic and highly scalable. Instead of building monolithic, self-contained LLMs for every micro-problem, it advocates for integrating smaller, specialized models or decision trees that interact with existing ERP, inventory management, and sourcing systems. For sectors like agri-food and manufacturing, this architecture minimizes data dependency risk while maximizing operational intelligence. This is the 'last mile' of adoption: moving past proof-of-concept demos to robust, profitable system integration.