Work
Case study · Vision AI

Inspection that runs on the line.

A food producer sampled 2% of production by hand and still missed defects. We put vision on the line — every unit inspected, alerts in under a second, all on their own hardware.

SectorFood & beverage
EngagementFeasibility + build
Timeline12 weeks to the line
Project film
01
At a glance
100%
of production inspected — up from 2% sampling
1s
from camera to operational alert
24
coverage without adding headcount
0
cloud round-trips — runs on-prem, on the line
02
Challenge
The problem

Sampling missed what mattered.

Manual QC could only sample a fraction of production, so defects surfaced where they cost the most — at the customer. The risk was carried by everyone downstream.

Staffing full inspection was impossible, and the cameras already on the line recorded everything while nobody watched.

  • 2% sampling, 100% of the liability
  • Defects found downstream, not on the line
  • Footage existed — the signal didn’t
03
Solution
What we built

Eyes on every unit, on-prem.

We started with feasibility on their real footage — before any commitment. Then detection was tuned to their actual conditions: the lighting, the speeds, the product variants.

The system runs at the edge, on hardware on the line, wired into alerts and the stop system. Accuracy is monitored and models retrain as conditions drift.

  • Feasibility proven on real footage before commitment
  • Edge hardware on the line — no cloud dependency
  • Alerts wired into the line and the dashboards
04
Result
The outcome

Every unit, every shift.

Inspection went from 2% sampling to every unit checked, alerts reach the operator in under a second, and defects are caught before packing instead of after delivery.

“We see the defect before the operator can blink.”

Background

Why “just add a camera” wasn’t enough.

The line already had cameras — recording for compliance, watched by nobody. The gap wasn’t hardware; it was a model tuned to their products and a path from detection to action the operators could trust.

Approach

How the twelve weeks ran.

Weeks 1–3 — Feasibility. We tested detection on their historical footage and agreed the accuracy bar before any build began.

Weeks 4–9 — Build & tune. Edge deployment on the line, detection tuned through the real variants and lighting shifts, validated shift by shift.

Weeks 10–12 — Wire & handover. Alerts into the line and dashboards, drift monitoring in place, operators trained on the override flow.

What's next

From one line to the plant.

The second line is being commissioned now — same models, new angles. Because the system runs at the edge and is monitored for drift, scaling is configuration, not another project.

Have a challenge like this?

Put eyes on the line.

Book a diagnostic