Open-source workplace safety
Catch PPE violations before they become incidents.
Upload a worksite photo and get violation detection, a GradCAM and SHAP explanation, an OSHA-grounded incident report, and a 7-day compliance forecast. Free and self-hostable.
Live detection
See what the model sees.
SafetyVision scans any worksite photo, finds each worker, and checks them against the PPE the model is trained to spot — hard hats, vests, masks, gloves, harnesses. Every item gets a labelled box and a confidence score, and anything missing is flagged in red, like the unprotected worker in this shot.
- ✓ A labelled bounding box around every worker and piece of PPE
- ✓ A confidence score on every single detection
- ✓ Missing gear flagged in red and ranked by risk level
How it works
From a single photo to an audit-ready report.
Upload a frame
Drop in a worksite image. Detection runs on a fine-tuned YOLOv8s model, no GPU needed.
See why it flagged
Every detection ships a GradCAM heatmap and SHAP attribution, so the call is never a black box.
Get the report
An OSHA-grounded incident report and a 7-day compliance forecast, exportable as a PDF.
Explainable by default
Detection you can actually trust.
Not a black box. Every violation comes with the evidence behind it.
- ✓ GradCAM heatmap over the detection region
- ✓ SHAP pixel attribution for the flagged class
- ✓ Confidence, bounding box, and risk level on every call
Grounded reports
Reports that cite the actual regulation.
Multimodal Gemini reads the annotated frame and writes the report, grounded in real OSHA 29 CFR text via RAG.
- ✓ Cites 29 CFR 1910 / 1926 by section
- ✓ Corrective actions and follow-up timeline
- ✓ One-click PDF with a human-review footer
Compliance forecasting
See the trend before the incident.
Prophet projects each violation type seven days out with an 80% confidence band, so risk surfaces early.
- ✓ Per-type 7-day forecast with uncertainty band
- ✓ Prophet vs SARIMA baseline
- ✓ Reads your own violation history
