Home/Solutions/Industrial Vision QA

Edge AI · NVIDIA Jetson

Inspect every part. At line speed.

AI visual inspection that runs on NVIDIA Jetson, right beside the camera. Find scratches, dents, missing components and misprints in milliseconds, and reject bad parts before they leave the line. No video has to leave your plant.

NVIDIA Jetson OrinMetropolisDeepStreamTAO ToolkitTensorRTGigE / USB3 VisionPLC · OPC UA · MQTT
Station QA-02 · camera 1 · model pcb-defects v7Simulated
Inspected0
Pass0
Rejected0
Inference-- ms
Line speed

What we build

Inspection that adapts to your product, not the other way round.

Rule-based machine vision breaks when lighting, materials or suppliers change. Deep-learning inspection learns what good looks like, and keeps learning on your line.

Surface defect detection

Scratches, dents, cracks, stains, voids and coating flaws on metal, plastic, glass, textiles and painted parts.

Assembly verification

Missing, misplaced or wrongly oriented components, connectors and screws. PCB inspection for solder and placement defects.

OCR and code reading

Lot and date codes, labels, serials, barcodes and DataMatrix codes read and verified against your MES records.

Measurement and gauging

Dimensions, gaps, fill levels and alignment checked with calibrated cameras, combined with AI for the hard cases.

Anomaly detection

Train mostly on good parts when defect samples are rare. The model flags anything that does not look normal.

Safety and zone monitoring

PPE checks, restricted zones and person-near-machine alerts using the same edge AI hardware and pipeline.

From camera to reject gate

One pipeline, built on the NVIDIA edge AI stack.

01 · Capture

Optics and triggers

Industrial cameras, lenses and lighting chosen for the defect. Image capture triggered by encoder or PLC.

02 · Train

Models with TAO

Labelled and synthetic defect images, fine-tuned from NVIDIA pre-trained vision models with the TAO Toolkit.

03 · Optimize

TensorRT

Pruning, distillation and FP16 or INT8 quantization so the model fits the Jetson and the cycle time.

04 · Deploy

DeepStream on Jetson

Multi-camera inference at the edge. Pass or reject sent to the PLC over digital I/O, Modbus or OPC UA.

05 · Improve

Close the loop

Uncertain images go to review, models retrain, and new versions roll out over the air through the ARP Device Cloud.

Why the edge

Decisions inside the cycle time.

A reject gate cannot wait for a round trip to the cloud. Running inference on NVIDIA Jetson next to the camera keeps decisions fast, private and available.

  • Millisecond decisions that keep pace with the conveyor
  • Video stays on site. Only results and sample images are uploaded
  • Keeps inspecting when the internet connection drops
  • Every station on the same model version, updated over the air

Hardware sized to the line

  • Jetson Orin NanoCompact smart-camera stations with one or two cameras and focused defect classes.
  • Jetson Orin NXMulti-camera stations, higher resolution and several models running at once.
  • Jetson AGX Orin and Jetson ThorHigh-throughput lines, many streams, and vision-language models for open-ended inspection.
  • Industrial enclosures and I/OFanless, rugged carrier boards with isolated I/O, PoE cameras and panel or DIN-rail mounting.

Where it works

Quality problems we are built to solve.

Electronics and PCBs

Component presence and polarity, solder bridges, tombstoning, and conformal-coating coverage.

AOI · assembly

Automotive components

Machined, cast and moulded parts checked for surface flaws, burrs, porosity and missing features.

Surface · dimensional

Pharma and packaging

Label, seal, cap and fill-level checks with code verification for traceability.

OCR · DataMatrix

Food and FMCG

Foreign objects, shape and colour grading, pack integrity and print quality.

Grading · integrity

Textiles and sheet goods

Continuous web inspection for weaving faults, stains, holes and pattern drift.

Line-scan · anomaly

Metals and castings

Cracks, cold shuts, scale and coating defects on hot or reflective surfaces.

Rugged optics

How we engage

Prove it on your parts first.

1

Feasibility

Send us good and defective samples. We test optics and models and report what is detectable, and how reliably.

2

Pilot station

One station on your line, running alongside manual inspection so results can be compared.

3

Production rollout

PLC integration, operator screens, MES reporting and validation documents for each station.

4

Continuous learning

Review queues, retraining and over-the-air model updates keep accuracy high as products change.

Have a defect your team keeps missing?

Send us a few samples. We will tell you whether AI can find it, and what a station would need.