Datavengers

Inspiration Every year, thousands of workers suffer preventable hand injuries in manufacturing environments. We were inspired by the potential of AI to go beyond historical reporting and offer real-time, actionable insights that empower safety teams to proactively reduce workplace risks.

What it does AI-Driven Insights for Injury Prevention uses large language models to analyze thousands of OSHA severe injury reports, classifying whether each incident was caused by human error or machine failure. It visualizes key trends and helps organizations identify the root causes of hand injuries—enabling smarter awareness campaigns and more targeted safety interventions.

How we built it We cleaned and filtered OSHA’s severe injury dataset to isolate hand-related incidents in the manufacturing sector. We then used OpenAI’s GPT-4o to process injury narratives in batches, labeling them based on causality (human vs machine). These insights were integrated into a pandas-based pipeline and visualized using Matplotlib. We tracked cost per API call and handled the model's outputs programmatically.

Challenges we ran into Creating clean, consistent labels from unstructured injury descriptions

Preventing API rate limits and ensuring error handling in long batch runs

Designing prompts that return strictly formatted outputs (1s and 0s)

Balancing model cost with coverage across 16,000+ records

Accomplishments that we're proud of Successfully labeled all hand injuries across a large dataset using LLMs

Created a reusable, cost-tracked, scalable pipeline

Designed targeted visualizations that speak directly to industry safety teams

Identified how human decision-making contributes to workplace injury patterns

What we learned How to integrate LLMs with structured data analysis at scale

The power of prompt engineering for consistent, machine-readable outputs

How to balance model cost, batch size, and real-world usability

That data storytelling can directly influence safety outcomes in industry

What's next for AI-Driven Insights for Injury Prevention Extend the model to classify types of human error (e.g., procedural bypass, distraction)

Incorporate time-series and location-based clustering for site-level safety diagnostics

Build an interactive dashboard (e.g., Streamlit) for clients to explore insights

Expand beyond hand injuries to cover other high-impact injury types

Partner with safety consultancies or industrial clients for real-world deployment

Built With

• amazon-web-services
• api
• openai
• python
• snowflake