The engine: real-time vision on real job sites.
This page explains what our models see, how they decide, and exactly how AWS and NVIDIA technology is used, not name-dropped.
Live frames from existing site cameras
Standard IP camera streams in varied light and weather: dust, rain, glare, night work under floodlights. We build and label our violation dataset from footage like this, because models trained on clean stock imagery fall apart on a real site.
The violations behind most injuries
PPE compliance per person per frame: hard hats and hi-vis vests. Workers inside exclusion zones near cranes, trenches, and heavy equipment. Blocked emergency exits. Each detection class exists because it maps to a documented cause of site injuries.
Persistence thresholds, not single frames
Detection and tracking models classify PPE and follow worker positions in real time. An alert fires only when a violation persists past a set threshold. One odd frame is noise; six seconds inside a crane radius is a hazard.
A feedback loop per site
Every confirmed and dismissed alert becomes labeled training data. The models sharpen for each site's specific cameras, mounting angles, and conditions, so coverage gets better the longer Sentrywork runs.
How we use AWS and NVIDIA: explained, not name-dropped.
How we use AWS
- Amazon Kinesis Video Streams ingests live camera feeds from connected sites.
- Amazon S3 stores confirmed alert clips under each customer's retention policy.
- AWS Lambda and Amazon ECS run alert orchestration: thresholds, routing, and notifications.
- Amazon RDS holds violation and compliance records that back the dashboard and exports.
- Amazon CloudFront serves the dashboard and this site; Amazon CloudWatch monitors uptime and stream health.
- Amazon Bedrock is planned for future incident summarisation across a site's alert history.
How we use NVIDIA
- NVIDIA GPUs train our PPE and proximity detection models on labeled site footage.
- TensorRT optimises those models and Triton Inference Server serves them for real-time, frame-by-frame inference at the rates safety detection requires.
- NVIDIA Jetson edge devices are our path for low-bandwidth sites: on-site detection without streaming raw video off site.
AWS and NVIDIA are trademarks of their respective owners. Sentrywork is not affiliated with or endorsed by them.
Three curves just crossed.
Models finally handle site footage
Vision models are now accurate on cluttered, low-light, weather-beaten footage: the kind real sites produce, not the kind demo reels use.
GPU inference is affordable to run continuously
Watching every frame of every feed used to be economically absurd. Optimised inference on modern GPUs makes continuous monitoring a subscription, not a data-centre project.
Insurers want proactive data
Carriers and clients are starting to ask for evidence of proactive safety monitoring, not just after-the-fact incident reports. Sites that can show it have an edge.
Who it's for, and where we honestly are.
Market
- Target users: general contractors, industrial facility operators, and safety consultancies.
- Launch markets: [confirm launch geographies]
- Revenue model: a per-camera or per-site monthly subscription; see pricing.
Traction, stated honestly
- MVP PPE detection in development
- Piloting on a small number of active sites
- Building the labeled violation dataset from real footage
- Waitlist open for contractors and operators
Where this is going.
| Milestone | What it delivers | Status |
|---|---|---|
| MVP PPE detection | Hard hat and hi-vis vest checks on live feeds, alerts with clips. | In development |
| Pilot & dataset build | First sites live; labeled violation dataset grows from real confirmations. | In progress |
| Zone & proximity alerts | Exclusion zones around cranes, trenches, and vehicle routes. | Next |
| Blocked exit detection | Alerts when emergency exits and escape routes are obstructed. | Next |
| Edge & offline processing | NVIDIA Jetson devices for low-bandwidth sites; no raw video leaves site. | Planned |
| New market expansion | Beyond launch geographies and into adjacent industrial verticals. | Planned |
See it on your cameras.
Book a walkthrough and we will show you what Sentrywork would catch on a site like yours.