Background: Challenges in Safety Reporting

Safety reporting is critical in construction, yet a global civil engineering company operating in Northern Africa faced significant barriers in collecting and acting on safety observations across their multilingual workforce and geographies. Managed and operated primarily in English, the company encountered operational challenges bridging language barriers on the ground. The existing process relied on paper forms filled out after shifts, which posed multiple challenges:

• Language Diversity: Workers spoke a wide range of languages, necessitating translation.

• Low Literacy Levels: Many workers struggled with formal reporting, limiting detail and participation.

• Manual Bottlenecks: Translators and reviewers spent hours categorizing reports and reclassifying them when errors were identified.

• Delayed Insights: The lag in processing reports delayed corrective actions and limited data-driven safety improvements.

The company sought a digital solution to streamline reporting, overcome language barriers, and improve data quality for better decision-making.

Solution: AI-Powered Safety Reporting with Fine-Tuned GPT-4o-Mini

AECFoundry implemented a comprehensive solution leveraging advanced AI and low-code tools to transform the company’s safety operations.

Fine-Tuned Language Model for Safety Classification

Using a dataset of approximately 12,000 historical safety observations, GPT-4o-mini was fine-tuned to classify new reports with an accuracy of 96%, compared to a baseline accuracy of GPT-4o-mini of 47%.

• Classify safety issues according to 25 predefined categories.

• Align with the company’s taxonomy and terminology.

• Handle multilingual inputs with high accuracy.

The model was trained to handle ambiguities in natural language and provide consistent categorization even in varied contexts.

Streamlined Multi-Lingual Input

Workers now report safety issues digitally using their native language. The system:

• Automatically detects the input language and translates submissions into English if necessary.

• Supports natural language reporting, eliminating the need for rigid form structures.

• Reduces the input process to two fields: location and the safety report.

Automated Categorization and Enrichment

When a user submits a report:

• The system classifies and categorizes the report in real-time.

• A priority level is assigned based on the report’s urgency and content.

• Sentiment analysis evaluates tone, while user-defined tags are added for enhanced reporting.

User-Friendly Interface and Reporting Dashboards

Airtable was integrated as a low-code platform to:

• Capture safety reports via a simple digital form.

• Visualize trends through intuitive dashboards.

• Manage workflows for tracking resolution status and generating reports.

  
  

Benefits: Real-World Impact of the Solution

1. Improved Accessibility: Workers of all literacy levels and language backgrounds can now contribute to safety reporting, increasing participation by 40%.

2. Operational Efficiency: Automated classification and translation reduced processing times by 10x, saving hundreds of hours monthly.

3. Data Accuracy and Insights: The fine-tuned model improved categorization accuracy by 5x, ensuring consistent reporting aligned with company standards.

4. Proactive Safety Culture: Real-time categorization and prioritization allow supervisors to address critical issues immediately, reducing incident response times.

5. Scalability and Adaptability: The solution’s modular design enables easy updates to tags, categories, and workflows, supporting future needs.

   
   

Technical Highlights

• Airtable: Low-code interface for digital forms, data storage, and dashboard visualization that the safety team can extend and adapt in the future.

• Azure AI: Secure training and deployment of the fine-tuned GPT-4o-mini model and safety report processing pipeline. Integrated translation services for multi-lingual support.

• Azure AI Translator: State of the art neural machine translator for language detection and translation in over 100 languages.

• GPT-4o-Mini: Fine-tuned with 12,000 labeled safety observations, achieving domain-specific accuracy and multilingual support.

  
  

Outcomes and Future Directions

The solution has modernized safety reporting for the client, creating a scalable system that enhances safety culture and operational efficiency. Key outcomes include:

• Increased Reporting Participation: 2x more reports submitted compared to the previous paper-based process.

• Faster Processing: Reports processed and categorized 70% faster.

• Data-Driven Insights: Enhanced tagging and categorization offer richer analytics for monthly and long-term safety planning.

Looking ahead, the company plans to:

• Add voice input to further simplify reporting for low-literacy workers.

• Implement real-time alerts for high-priority hazards.

• Integrate the system with task tracking tools for seamless issue management.

• Implement predictive analytics on collected data to identify at-risk projects based on sentiment, reporting activity, and risk profile.

  
  

Conclusion

By fine-tuning GPT-4o-mini and leveraging low-code tools, AECFoundry delivered a robust, scalable, and impactful solution to overcome language, literacy, and process challenges in construction safety reporting. This case study highlights how AI-driven innovation can bridge operational gaps, empower diverse workforces, and foster a proactive safety culture in complex industries.

Interested in transforming your Health Safety and Environment management and reporting processes with AI? Contact Us today to learn how we can help you achieve similar results.