Integrating pre-trained AI models into operational workflows: PPE detection for mine sites
Overview
Gulf Consulting was engaged to evaluate whether a pre-trained, cloud-hosted AI image classification model could be integrated into mine site operations to support workplace safety monitoring — without the cost and complexity of building a custom model. The project reflects a core capability of our practice: identifying where off-the-shelf pre-trained AI models can deliver measurable process improvements quickly, and where deeper investment is warranted.
Background
The engagement focused on a question our clients frequently raise: can a pre-trained model, activated as-is, deliver enough value to justify integration into an existing process? To test this, we selected the AWS Personal Protective Equipment (PPE) detection model and paired it with a lightweight web application. No fine-tuning was performed, and no retrieval-augmented generation (RAG) was applied. The intent was to isolate the usability and performance of the base model as an integration candidate.
Solution Delivered
We designed and built a web-based image interface allowing operational staff to upload photographs of workers on site. The AWS PPE detection model analysed each image, identifying head, face, and hand PPE per person, drawing bounding boxes over the detected items, and returning confidence scores for each detection. Users could then export results directly to a formatted PDF report suitable for compliance review and record-keeping. The end-to-end workflow — from image capture to reportable output — was operational within a short timeframe and required no model training.
Results achieved and takeaways
Testing revealed a clear performance profile for the base model:
- Detection of head and face PPE was consistently strong.
- Glove and hand detection was materially weaker, likely reflecting an imbalance in the underlying training data.
- Image quality and orientation had a significant effect on accuracy.
- Response speed was excellent, supporting near-real-time use.
- Confidence thresholds were configurable, allowing the workflow to be tuned to the client's tolerance for false positives and negatives.
The project reinforced several principles that shape how Gulf Consulting approaches AI integration for our clients:
- Cloud-hosted, pre-trained models are a low-cost, low-risk entry point for delivering measurable process improvements — a fast route to reduced cycle times and stronger precision and recall on well-supported detection classes.
- Fine-tuning through transfer learning is required to reach enterprise-grade accuracy, particularly for detection classes underrepresented in the base model.
- Complex workflows will often call for multiple models working in combination, rather than a single general-purpose model carrying the full load.