Overview
Punch and bend presses are some of the most dangerous machines on a factory floor. SafeHands is a computer vision safety system that watches the danger zone in real time and stops the machine the instant a hand or arm enters it. I built everything except the detection model itself — backend, frontend, camera integration, and the real-time pipeline connecting them. It won 1st place at AI+ Day 2022 and has run in production at Panasonic ever since, with zero recorded incidents.
The Problem
Traditional safety guards slow operators down, so operators disable them to hit production targets — trading safety for speed. Panasonic's case was harder still: their metal bending machinery used a non-standard configuration that existing safety hardware couldn't be adapted to. What was needed instead: a system that could tell a hand from a workpiece, in real time, fast enough to matter.
What I Built
A neural network (built by a teammate) recognizes hands and arms; everything around it — getting that detection to actually stop a machine in time — was mine.
- Built the full React.js frontend and the Flask backend from scratch.
- Used Server-Sent Events to stream detection state from model to interface in real time, so a stop event shows up the instant it happens on the floor.
- Integrated with Advantech's camera web API, building a custom interface that replaced their standalone camera app with just the controls our use case needed — including debounced calls and stepped zoom limits to keep the hardware from getting hammered with redundant requests.
- Built a 6-point ROI overlay on the live video feed so the AI engineer could define exactly where hand detection should trigger — no existing library supported this. Started in raw Canvas, which choked on re-renders, then rebuilt it in Konva.js for a smooth, responsive overlay.
- Streamlined deployment, cutting setup time per installation from a full day to 30 minutes.
Results
- 100% detection accuracy identifying limbs and triggering an automatic stop, even in setups where traditional safety guards had been disabled.
- 150ms response time from detection to shutdown — fast enough to act before contact.
- Deployed at Panasonic since 2022, solving a safety problem their existing equipment couldn't — with no reported accidents as of 2026.
- 1st place, AI+ Day 2022.