PPE compliance

Catch every missing hard hatbefore the shift ends.

Continuous, per-feed PPE verification across every area, every shift. No more quarterly photo-ops or clipboard walkarounds — just real-time evidence and a queue your supervisors actually clear.

Book a demoSee zone intrusion
Hard hatHi-vis vestSafety gogglesCut glovesHearing protectionFace shield
The problem

PPE compliance gets measured in the wrong places.

Manual inspection samples minutes from a week that's 10,080 minutes long. The missing coverage is where the incidents live.

Shift change
is when most PPE citations happen.
Walkarounds are weakest in the first 30 minutes and the last 30 minutes of every shift.
2.4×/wk
Average manual PPE walkarounds per area.
A supervisor sees roughly 0.1% of actual shift-hours in any given zone.
6×
Under-reporting of PPE-related near-misses.
Industry estimate — source available on request.
The solution

Detect, attribute, and log — every frame, every feed.

Ten ships-with PPE classes, rules framed in plain English, and an operator UX designed for people who wear steel-toes to work.

Rule framing

“Hard hat + hi-vis required in Area B from 06:00 to 22:00.”

Every rule reads like a sentence a safety manager would actually write. Behind the scenes it compiles to a deterministic evaluator with an audit trail. Change history is diff-able — you can prove to an auditor exactly which rule was active on the day of an incident.

area Area B +
shift 06:00 – 22:00 +
requires hard_hat, hi_vis_vest
→ alert on violation ≥ 4s dwell

Live feed wall

Up to 24 cameras at a glance. Violations highlight the offending worker with a class label.

Event queue

Triage, acknowledge, reassign, or mark false positive. Every action is logged with operator ID.

Weekly compliance PDF

Auto-generated, area-by-area breakdown. Drop it into your ISO 45001 audit pack unchanged.

Benchmarks

Technical measurements, not marketing copy.

Lab-measured figures on reference hardware. Pilot-stage deployment numbers will replace these once available.

< 180ms
P95 inference latency
NAO-A20, 1080p feed, lab-measured
12×
More coverage vs. manual walkarounds
Continuous vs. periodic spot-check
< 60s
Alert latency (rule match → Teams)
On-appliance, no cloud hop
3 wks
Typical time to full-site rollout
From PO to all areas live
Example rules

Three rules, copy-paste-ready.

The editor is visual — but rules export as YAML for version control and peer review.

Baseline · Area B day shiftyaml
rule: ppe-area-b-day
area: polygon("Area B - Assembly")
schedule: Mon-Sat 06:00-22:00
required_on: person
required_items:
  - hard_hat
  - hi_vis_vest
dwell_seconds: 4
on_violation:
  - notify: teams("#ehs-plant-04")
  - snapshot: s3://evidence/ppe/
  - escalate_after: 5m
Grinding cell · full kityaml
rule: ppe-grinding-cell
area: polygon("Grinding Cell 3")
schedule: always
required_on: person
required_items:
  - hard_hat
  - safety_goggles
  - hearing_protection
  - cut_gloves
dwell_seconds: 2
on_violation:
  - notify: teams("#grinding-ops")
  - plc_out: "bay3.warning_lamp"  # amber beacon
Visitor walkthroughyaml
rule: ppe-visitors
area: polygon("Visitor Corridor")
schedule: always
applies_to: person[role=visitor]
required_items:
  - hi_vis_vest
  - safety_goggles
on_violation:
  - notify: email("reception@site.com")
  - hold_badge: true
Model & data

The model that ships with every PPE release bundle.

Every release bundle carries a signed model card and datasheet. The cross-repo contract below is what the on-appliance inference stack enforces at load time.

Architecture
RF-DETR-Nano (transformer-based object detector)
Trained on public PPE datasets; fine-tuned per-customer on redacted site footage when needed.
ONNX signature
input: 1x3x640x640 · output: detection boxes + class probs
Frozen across releases — changing it requires a coordinated core-library update.
Class coverage
Hard hat, hi-vis vest, safety goggles, face shield, cut gloves, hearing protection, person, and four negative-case classes.
Exact index mapping lives in the release bundle's labels.txt.
Eval metrics (baseline)
— pilot-pending
Real per-class mAP@[.5:.95] published with each release once a pilot baseline exists.
Release cadence
Quarterly major, monthly minor for per-site fine-tunes.
Every release ships a model card (Gebru datasheet + Hugging Face model-card format).
Weight provenance
Allowlisted sources only — Apache-2.0 or MIT.
Training refuses to start if the base checkpoint is not in configs/weight_allowlist.yaml.
Integration

One detection, end-to-end, in under a second.

Events flow from the camera through the on-appliance event bus to your operator tooling. Everything stays on-premise until the last hop.

01
RTSP camera
Your existing cameras — no replacement, no new cabling.
02
On-appliance inference
RF-DETR runs on the NAO-A20 / A100 GPU — no frames leave the site.
03
Rule evaluator
Deterministic evaluator matches detection against area + schedule + required items.
04
MQTT event bus
Topic detections/{camera}/{rule} — mTLS authenticated, ACL-gated.
05
Teams / email / webhook
HMAC-signed outbound — or SIEM via syslog if preferred.
ROI estimate

What continuous monitoring replaces, in labour hours.

A conservative calculator. Adjust the inputs to your reality. The only number that appears is the one your inputs produce.

formula: (minutes × checks × days / 60) × €/h

Labour hours avoided per year
400 h
At your hourly cost
€18,000 / year

Incident-cost avoidance and regulatory-fine exposure are NOT included — those are customer-specific and the founder does not publish estimates without real pilot data.

Run this on your data

Every shift. Every area.

Show us a recorded clip of a shift change. We'll come back with a live PPE-compliance preview against your actual site.

Book a demoSee the platform