Build and own a Hazard Recognition Specialist Model.
Does this site image, video feed, or sensor reading show a safety hazard, and if so, what type and severity, requiring what response?
Overview
Reduces workplace injuries by detecting hazards before an incident occurs.
Who owns this decision
VP Environmental Health and Safety (EHS), Chief Safety Officer
Why owned AI may fit
Hazard patterns are specific to a site's layout, equipment, and operational history, and visual hazard detection requires domain-tuned computer vision rather than general image understanding.
What this Specialist Model actually does
Decision: Does this site image, video feed, or sensor reading show a safety hazard, and if so, what type and severity, requiring what response?
This is a larger system with 3 specialist model candidates and 1 supporting or external component.
Who uses this Specialist Model
- VP Environmental Health and Safety (EHS)
- Chief Safety Officer
- Director of Site Operations
- Head of Risk Management
Where the specialist model sits
This is a larger system with 3 specialist model candidates and 1 supporting or external component.
Specialist vision model
Specialist Model
Specialist vision model for real-time hazard detection and classification from camera feeds
Sensor-fusion model combining wearable/environmental sensor data
Specialist Model
Sensor-fusion model combining wearable/environmental sensor data with visual detections
Specialist LLM
Specialist Model
Specialist LLM for generating inspector-facing hazard reports and response recommendations
Real-time video and sensor ingestion pipeline
Supporting component
Real-time video and sensor ingestion pipeline with edge inference for low-latency alerting
How Oumi connects to your stack →
How to evaluate and improve it
Evaluation
- Lead time between detection and incident prevention
- Incident rate reduction vs. baseline
Feedback signals
- Safety inspector confirmation or dismissal of flagged hazard
- Actual incident occurrence following a hazard alert
- Near-miss report correlated with prior detection
- Worker/supervisor override or dispute of classification
- Regulatory citation outcome tied to a site area
- False-alarm rate from field verification
Model candidates to evaluate
- qwen3-vl
- modernbert-large
- qwen3.5-9b
These are candidates, not performance claims. Run them against your evaluation set.


