Safety incident classification
Summary
Incident reporting for a manufacturing safety team.
Operators described similar incidents in different ways, making reports hard to classify consistently against the company’s safety taxonomy.
A service that interprets incident descriptions, assigns structured categories and prepares reports using the client’s terminology. Safety specialists review and approve the output.
Tech Stack
- Python
- FastAPI
- Pydantic
- LangChain
- LangGraph
- Azure OpenAI
- PostgreSQL
- Docker
- Azure
- Langfuse
- OpenTelemetry
Tech Challenge
- High incident volume. A large number of safety incidents are recorded every day across the enterprise, making manual processing at scale impractical without automation.
- Inconsistent manual classification. Human operators classify incidents differently depending on experience and interpretation, introducing inconsistencies into safety records and hampering trend analysis.
- Alignment with internal standards. Incident descriptions must adhere to the company's own reporting rules and vocabulary. Generating text that is both accurate and compliant requires a deep understanding of internal policies.
- Significant human effort. Drafting rule-compliant incident descriptions from raw event data is a labour-intensive task that ties up Safety Team capacity that could be used for prevention and analysis.
- Need for efficiency and accuracy. The Safety Team required a solution that would improve throughput without sacrificing the quality or auditability of safety records.
Solution
- The classification engine analyses raw incident reports and proposes incident types using the client’s internal taxonomy.
- The description-generation module turns unstructured reports into structured drafts using the company’s reporting rules. Safety specialists review and approve them.
- Adapted the solution to the client's data, terminology, and reporting context so that classifications and generated text reflect the operational realities and language conventions of the manufacturing environment.
- Designed the pipeline for seamless integration into the existing safety management workflow, allowing the Safety Team to review and approve AI-generated outputs without disrupting established processes.
- Implemented quality and consistency controls to ensure that outputs are reproducible, auditable, and ready for regulatory or internal review without additional manual editing.
Impact
Outcome
A more consistent classification and documentation workflow for safety teams.
Safety specialists review and approve generated records. Classification and documentation support do not replace that review.
Published
