01
Sectors
02
AgentMesh
03
AgentNexus
04
Hardware
05
Why Unovie
06
Case study
07
FAQ
08
Pilot on your line →
Home/Research/Case study
Case study · Vision at the edge · Semiconductor

A thousand cameras, and every decision made on site.

A semiconductor manufacturer wanted cameras watching every chamber on its lines: 25 cameras per chamber, 40 chambers per site. Nobody can watch a thousand feeds, and sending them all to the cloud would have been slow, expensive and a risk for sensitive footage. They had only a couple of dozen labelled images to start from. We proved the approach in an MVP on their own footage, analysing more than 300 frames per second on site, then designed the site so the vision AI runs next to the cameras, every feed is checked as it arrives, and the video stays in the building.

25
cameras per chamber
40
chambers per site
1,000
camera feeds watched
300+
frames per second analysed on site in the MVP
01 · At a glance

The engagement.

ClientSemiconductor manufacturer (name withheld)
ProblemCameras on every chamber, but no way to watch or act on a thousand feeds in time
Starting pointAbout 25 labelled images and a few videos
Scope25 cameras per chamber · 40 chambers per site · about 21 TB of video a month
DeploymentOn site, next to the cameras, on NVIDIA Jetson hardware accelerators. No cloud in the decision path
EngagementMVP on the customer's own footage, then a site design sized from the MVP results
MVP result300+ frames per second analysed on site, on the customer's own footage
02 · The problem

Too many cameras, not enough eyes.

Nobody watches the feeds

Cameras get installed for safety and quality, then mostly record. Problems are found later, on a replay, after the parts are scrapped.

The cloud adds delay

A round trip to the cloud adds delay and depends on the network link. A decision that matters on the line can't wait for either.

Costs grow with every camera

Per-camera cloud pricing looks small at ten cameras. At a thousand, it becomes one of the biggest lines in the budget.

Models go stale

Lighting, fixtures and products change. A model set up once and left alone gets worse every month.

03 · What we built

Watch every feed, flag what matters.

Decides next to the camera

Each feed is checked as it arrives, on hardware in the customer's building. No cloud round trip, and it keeps working if the internet link drops.

on siteno round trip

Shows its proof

Every flag comes with the frame and a short note on what was seen, sent to the operator or engineer who owns that chamber.

proof frameright person

People decide

The system flags and asks. It never moves equipment. The customer's control systems and people decide what happens next.

alarm and requestno equipment control

Stays accurate

Starting from about 25 labelled images, we built the first models and look after them from there. Improvements are tested on the customer's own footage before they go live, and can be rolled back.

tested on your datareversible
04 · What the site gets

Built for the factory floor.

CamerasRugged industrial cameras, made in the USA and NDAA compliant, with depth sensing and night vision. Existing cameras can be used too.
On-site AI hardwareUnovie's custom-built NVIDIA Jetson hardware accelerators, sized to the site's camera count. Keeps running if one unit fails.
Video stays on siteFootage is analysed and kept on site. Archiving it anywhere else is the customer's choice, not a requirement.
Connected to existing systemsFlags go to the existing alarms, MES, quality and work-order systems.
Looked afterModel updates, hardware support and monitoring included. Buy the hardware or take it as a service.
05 · What carries over

The same engine, on other lines.

This project was in a chip fab. The same engine is set up for other lines during a pilot, on their own footage.

Test stands

Watches a test run alongside the sensors and flags anything unusual to the test conductor as it happens.

space & aerospace

Battery welds and tabs

Checks welds and tab alignment in line, together with the process signals for the same cell.

battery

Robot cells

Spots misplaced parts and drifting motions before they turn into scrap or a stopped cell.

robotics

Conveyors and docks

Watches jams, spills and blocked lanes on conveyors and loading docks and alerts the shift lead.

material handling
06 · How we start

Small, then site-wide.

StepWhat happens
Site auditNo charge, 1–2 days. An engineer walks your floor, looks at your cameras and network, and picks the first tools to watch.
6-week pilotA handful of cameras on one tool, on your own footage, with a results report you can check.
ProductionRoll out chamber by chamber, then site by site, in about a quarter.

Customer details are withheld under confidentiality. The 300+ frames per second was measured in the MVP on the customer's own footage. Camera, chamber and video figures are the customer's site requirements.

Cameras already on the line?

See how our systems can help your line.

Book a short live demo. We'll show the vision engine checking a camera feed on site, and talk through what it would watch on your floor.