About the company
This client is an established name in consumer food production. Specialising in baking mixes for both retail and professional markets, the company combines deep product heritage with an ongoing drive to modernise its operations. Quality and consistency are central to its brand promise, making any source of production variability a direct business concern.
The challenge: color is more complex than it looks
Reproducing color accurately in baked goods is surprisingly difficult. The final color of a product after baking depends on a wide range of variables: the base recipe, the specific dyes used, their concentrations. For decades, getting this right relied on the intuition and experience of a small number of skilled employees, who worked through the problem by trial and error.
This approach was slow, costly in materials, and difficult to scale. When experienced staff were unavailable, consistency suffered. The company needed a more systematic solution: a way to encode that expertise into a reproducible, data-driven process.
The central question put to DataNorth: can an AI model, trained on historical recipe and color data, predict which dyes and quantities are needed to achieve a target color?
The solution: three complementary aI models
After an initial analysis phase, DataNorth developed a system of three models, each addressing a different part of the formulation problem.
- Forward model (ingredients to color) Given a base recipe and a set of added dyes, this model predicts the resulting color, expressed as Lab values. This creates a digital simulation of the baking process, without needing to run a physical batch.
- Inverse model (target color to dye selection) This reverses the question: given a desired color, which dyes should be used? This is inherently harder to solve, because the search space is larger. Different coloring ingredients can be combined in different compositions, and in different quantities, and multiple dye combinations can produce the same visual result. DataNorth built an optimisation model that generates and ranks multiple candidate solutions by accuracy. The optimisation loop uses the forward model to evaluate each candidate and steer the search in the right direction.
- Quantity model (dye selection to optimal amounts) This model takes the base recipe, the selected dyes, and the target color as input, and returns only the quantities needed. By narrowing the search space to amounts rather than both selection and amounts simultaneously, this model achieves higher precision on the final output.
Data analysis
The training dataset consisted of many historical recipes, each containing base ingredients, dye specifications, and measured color codes. Before model development, DataNorth conducted a structured data analysis to map the dataset: which colors appear in the data, how frequently each dye is used, in what concentrations, and which dye combinations tend to appear together. This analysis informed both model architecture and training strategy.
The result
The system transforms a process that previously depended on individual expertise and physical trial-and-error into a structured, repeatable workflow. Formulation time is reduced, material waste decreases, and the company is no longer dependent on a handful of experienced individuals to maintain color consistency across its product range. The models encode institutional knowledge in a form that can be used, validated, and improved over time.
