Skip to content
IEMI Labs
All product designs
VisionDesign complete (stages 1–6), not built

QualityCheck Bench Station

Teach-by-example visual inspection for small manufacturers, proven on your parts first

Checks every part at the bench or on a slow conveyor in under half a second, shows PASS, FAIL or UNSURE with the image, drives a reject output and logs each lot, after being taught with 20 good and 20 bad parts.

Who buys it
Owner or quality manager at a 10–200-person injection moulder, contract electronics assembler or small packer.
Why it pays
About 0.5 FTE of manual final inspection ($20–25k/yr) reduced to spot checks, plus fewer customer chargebacks (est.); payback under 3 months on a $2,900 station (est.).
Target price
$1,500 feasibility study (credited) + $2,900 station + optional $49/month (target)
First-minute wow
Place 20 good and 20 bad parts in the nest, and ten minutes later a red FAIL with a heat map shows exactly where the short shot is.
See the CAD designConcept page
QualityCheck Bench Station concept: the idea and where each part goes
The concept. The CAD design is below, under Printable CAD.

What the research found

And how this design answers it

  • Every line is custom → a fixed light box and part nest standardise 80% of jobs; the rest is a fixed-price $1,500 study with a GO/NO-GO
  • Overview OV20i lists at $9,450 and Cognex In-Sight 2800 is about $4,500 → $2,900 installed with lighting, fixture and isolated I/O included
  • 20/20 teaching is not proof → UNSURE band, human review queue and a 300-part acceptance run before go-live
  • Hailo compiling needs Python on x86 → precompiled backbone with tile kNN and a logistic head in Node.js, so no per-customer compile
  • Concept page over-promised 4K/IP65/10 parts per second → honest v1 limits: ≤2 parts/s, bench/indoor, 4 in / 4 out 24 V

Use case

Short shots caught at the press

Press operatorQC leadOwner

  1. 01Trigger: every 30 s the operator puts a moulded housing in the nest and taps the foot switch
  2. 02Detection: the ring strobes, the camera captures, and 16 tiles are scored against the taught good parts in about 300 ms
  3. 03Alert: FAIL turns the light red and fires the reject output; UNSURE parts go to a review tray
  4. 04Action: the QC lead labels the review tray at shift end, the model retrains in seconds, and the lot report is filed

Outcome: Every part checked with an image record per lot; about $20–25k/yr of manual inspection reduced to spot checks (est.)

How it works

Architecture

  1. 01

    Sense: Raspberry Pi Global Shutter or HQ camera in a white light box with a strobed 24-LED ring and a printed part nest

  2. 02

    Decide: Edge Hub (Pi 5 + Hailo-8L) runs fiducial registration, tile embeddings, kNN anomaly scoring and classic tools

  3. 03

    Act: isolated I/O pod (ESP32-S3) takes the trigger and drives reject and stack-light outputs, failing to reject if the hub goes quiet

  4. 04

    Tell: operator kiosk screen, optional cloud dashboard with lot reports and yield-drop alerts; images stay on site

Hardware

Real parts, buyable today

Bill of materials: about $620 per station (est.).

BlockPartWhy this part
ComputeRaspberry Pi 5 8GB + AI HAT+ 13 TOPS (Hailo-8L)Shared IEMI Edge Hub; one camera at ≤2 parts/s
CameraRaspberry Pi Global Shutter Camera (IMX296, C-mount)$63.75; no motion blur on a conveyor
Lighting24-LED ring PCB with Diodes AL8860 driverStrobed, diffuse, repeatable inside a white box
I/OESP32-S3 + Toshiba TLP293-4 + TLP241A4 isolated 24 V inputs and outputs, deterministic reject timing
PowerMean Well GST60A24 listed adapterNo mains inside anything IEMI builds

Mechanical

Printable CAD

8 parts, generated from code and checked as closed solids. Pick a part to inspect it, or download the STL files to print.

See the CAD parts render
QualityCheck Bench Station printable parts, rendered from the CAD
Render of the printable parts from the actual CAD. Not yet printed or photographed.

Loading the CAD…

Real CAD · drag to rotate · scroll to zoomReal CAD · swipe sideways

Open risks we are working on

  • Feasibility studies may not convert to station sales (need ≥40% by study 10)
  • Running Hailo inference from Node and choosing a precompiled backbone not yet proven
  • Tiny defects under 4 pixels are out of reach; the optics rule must be enforced in every study
  • Support load per station against a $49/month plan

Design stage only: nothing on this page is for sale. Specifications and prices are targets.