Worker AI Safety

Inspiration

We kept coming back to one uncomfortable fact: most workplace injuries aren’t caused by people who don’t care. They happen on a busy day, when someone forgets a helmet or skips a vest, and nobody happens to be looking.

A safety manager can’t be everywhere. On a construction site, they walk the zones once in a while and hope for the best, which leaves almost every hour of the week unwatched. And when an injury does happen, it can cost a company tens of thousands of euros. The real cost is that someone doesn’t go home the way they came in.

Most sites already have cameras. We wondered: what if those cameras could quietly look out for people, so someone is always paying attention?

What it does

WorkerGuard AI watches camera footage and notices when a worker is missing protective equipment, like a helmet, a safety vest. For now it uses live camera and recorded video for demo goal.

When it spots one, it sends an alert right away and shows it on a live dashboard, so the safety team knows what’s happening now. Every incident is also saved to a log with the time and details, so the team can look back, see where problems keep happening, and talk to people about it later.

It’s not meant to catch anyone out. It’s there to give safety teams a second pair of eyes.

How we built it

We used computer vision to find people in the video and check what they’re wearing. Detecting a helmet is easy. The harder part was deciding what counts as a problem: a worker with no helmet is a violation, while a helmet sitting on a table isn’t.

We then connected that logic to a dashboard showing live alerts, recent incidents, and the full log. We spent a lot of time on how it looks and reads, because a safety tool only works if someone can understand an alert at a glance and act on it.

Challenges we ran into

The biggest one was the gap between “the model sees something” and “a person can do something about it.” Turning detections into clear, trustworthy alerts took far more thought than we expected.

Alert fatigue was a close second. If a system cries wolf too often, people stop listening, so we had to find a balance between catching real problems and not drowning people in notifications. Real footage also doesn’t look like clean test videos. Bad lighting, crowded scenes, and people half-hidden behind equipment all make detection harder.

Accomplishments that we’re proud of

We’re proud that this is a whole, working loop and not just a model: it detects, alerts, displays, and records. We started with an idea about making workplaces safer and ended with something a safety team could actually use.

What we learned

We went in thinking the AI model was the main event. It turned out to be only one part. How alerts are worded, how incidents are shown, and how records are kept matter just as much. We also learned that trust is part of the product. Monitoring only works if people feel it’s there to protect them, not to watch them.

What’s next for WorkerGuard AI

First, we want to make it better at the basics: more accurate, fewer false alarms, and tested in tougher conditions like low light and crowded sites. We’d also like it to recognize more equipment, like goggles and harnesses, and send alerts where people already are, such as their phone, Teams, or Slack.

After that, we’d like to add danger-zone alerts, fall detection, and simple reports that show where and when risks pile up.

Privacy will stay at the center of all of it. We want the detection to run on-site so raw video never leaves the premises, with faces blurred and a focus on safety behavior rather than on individuals.

In the long run, we want WorkerGuard AI to be the quiet helper that makes sure everyone gets home safe at the end of the shift.

Your cameras already see everything. We help them look out for people.

Built With

• next
• python
• yolo