Predictive Quality Analytics for a Leading Irish Food Manufacturer

The Challenge

A leading Irish food manufacturer had invested significantly in collecting detailed quality control data across their production process. For every product batch, they recorded over 150 attributes — raw material specifications, processing parameters, environmental conditions, and time-temperature profiles tracked across multiple stages of production.

They also had something most manufacturers don't: formal quality assessments from trained evaluation panels, scoring finished products across multiple quality dimensions. Combined with their HACCP process controls and raw material specifications, there was a rich dataset — but no way to connect production inputs to quality outcomes at scale.

The data sat in spreadsheets. The quality team knew their products varied, but they couldn't explain why. Some batches consistently scored well. Others underperformed. The patterns weren't visible in Excel.

The core questions were:

This wasn't an AI-for-the-sake-of-AI project. It was a commercial question: can we use data we already collect to make better production decisions?

Our Approach

Barry Gough, our COO, led the project with support from our specialist data science team — experienced ML engineers and applied statisticians with backgrounds in production systems and explainable AI.

We took a systematic process analytics approach combining statistical rigour with modern machine learning:

1

Data Preparation and Exploration

Week 1

The raw data covered 350+ product records with 150+ attributes each. Before any modelling, we spent the first week on data quality: auditing for missing values, outliers, and inconsistent coding. We restructured the data into a clean, reusable schema designed to accommodate future production runs.

A critical step was feature engineering from the time-temperature profiles. Raw sensor curves can't be fed directly into models, so we extracted meaningful features: initial values, rates of change, time to key thresholds, area under the curve, and inflection points. These engineered features captured the process dynamics that matter for quality outcomes.

2

Statistical Analysis

Week 2

We generated comprehensive R² correlation matrices showing how each production attribute — individually and in combination — relates to each quality dimension. This highlighted which measurements actually matter and which add no predictive value.

We also ran reverse analysis: taking the top and bottom quartile products and working backwards to identify what characterises the best versus the worst performers. This often reveals patterns that forward modelling misses.

3

Machine Learning Models

Week 3

For the AI implementation phase, we evaluated multiple approaches to find what works best for this specific data: gradient boosting (XGBoost), random forests, support vector machines, and explainable boosting machines. We tested both regression (predicting exact scores) and classification (predicting quality grades) to determine which gives the most practical results.

All validation was performed at the batch level using proper cross-validation — ensuring honest accuracy estimates that reflect real-world performance, not overfitting to the training data.

We used SHAP analysis to make every prediction explainable: not just what the model predicts, but why.

4

Deliverables and Prototype

Week 4

We packaged everything into clear, actionable outputs: a comprehensive technical report, an executive summary for senior stakeholders, and a working prototype tool that lets the quality team input production attributes and see predicted quality scores with explanations.

What We Delivered

The Outcome

The analysis confirmed that a subset of production attributes — significantly fewer than the 150+ being measured — drives the majority of quality variation. The client now knows which measurements matter and which they could potentially reduce or eliminate.

"This was the first time the quality team could see — in data — what they'd always suspected about which attributes actually matter. The ranked list gave them something concrete to act on."— Paraphrased from project debrief

The predictive models achieved strong predictive accuracy for the primary quality dimensions and moderate accuracy for secondary dimensions — giving the production team a tool they can use alongside their existing processes. The best-performing models correctly distinguished top-quartile from bottom-quartile products in the majority of cross-validation runs.

The engagement was structured as a fixed-price Phase 1 project. An optional Phase 2 was scoped for production tooling and a sampling protocol — designed to be informed by Phase 1's findings about which attributes prove most predictive.

The client used the Phase 1 findings to inform their quality improvement programme and has scoped Phase 2 for production-grade tooling.

Results vary by business. Figures shown are measured and estimated during delivery.