Pre-launch · design-partner pilots open for 2027Synthetic defects · edge inspection · NVIDIA stackApply for the pilot
CAM 03 · Line A · Station 2Model AX-204 adapter v7120 parts/minSimulated example

Visual inspection for manufacturing

Train on defects your line has not made yet

Viseena renders labeled defect images from your CAD, trains a model on them, and runs it on edge hardware at the station. Real frames come back so the model keeps up with every new SKU.

Inspected0
Pass0
Reject0
Review0

Simulated animation, not footage or results from a customer line. Viseena is pre-launch.

Design targets

What we are building toward

These are the targets the platform is designed around. We will publish measured results from pilots, not before.

Target24h

From CAD and a defect list to a validated model for a new SKU.

Target<50ms

Per frame on the edge node, across four camera streams.

Design property100%

Of synthetic frames arrive with pixel masks, boxes and classes.

Target0

Real examples of a rare defect needed before training starts.

The problem

The defects that matter most are the ones you rarely see

Rule-based machine vision has to be reconfigured whenever a part, material or light changes. Deep-learning inspection needs thousands of labeled examples per defect class, but real defects are rare by design and slow to collect.

So lines either over-reject and scrap good parts, under-detect and ship bad ones, or stall for weeks retraining on every new SKU.

  • Rare defects get too few training examples
  • Hand-labeling stalls every changeover
  • Lighting and material changes break rule sets
  • Escapes surface as returns and recalls
Defect frequencyExample

Why waiting for real defects fails

ScratchCommon
DentOccasional
PorosityRare
CrackRare
Missing featureVery rare
Illustrative distribution. On most lines the costliest defects are the least frequent.
Synthetic data

Defects, rendered and labeled before they happen

Switch between the image a model trains on, its pixel label and its bounding box. Every synthetic frame carries all three.

Illustrative tiles drawn procedurally on this page to show the idea. Synvex renders from your CAD and materials with NVIDIA Omniverse Replicator.

The platform

Five products, one inspection loop

Each product does one job. Together they move a defect from CAD to a decision at the station and back into the next model.

How it works

From CAD to the line in four stations

  1. ST-01Hours

    Generate

    Upload CAD, materials and the defect classes you care about. Synvex renders labeled variants across lighting, pose and texture.

  2. ST-02Hours

    Train

    Coreon fine-tunes a per-SKU adapter on the synthetic set and aligns it to a small batch of real frames.

  3. ST-03One shift

    Deploy

    Veridex runs the model on a Jetson or IGX node beside the cameras and diverts rejects through your PLC.

  4. ST-04Continuous

    Learn

    Uncertain frames return for review. Confirmed frames retrain the adapter and widen the defect taxonomy.

Timings are design targets for a part with existing CAD and an installed camera station.

Who it is for

Built for the people who own quality on the line

  1. 01

    VP Quality and plant managers

    Cut scrap and escapes without adding inspectors, and get an audit trail for every decision.

    Quality leaders
  2. 02

    Industrial and automation engineers

    Commission stations in a twin, then deploy models without writing rule sets.

    Engineering
  3. 03

    Integrators and machine builders

    Embed a self-adapting inspection runtime in the cells and stations you already build.

    Channel
Built on NVIDIA

GPU work at every step, one API over all of it

Rendering, training and edge inference are all GPU workloads. Every step you take in the Viseena app is also a REST call, so integrators can script it from their own tooling.

Omniverse ReplicatorDomain-randomized, auto-labeled synthetic defect scenes
TAO ToolkitTransfer learning and per-SKU adapters
DGX / HGXSynthetic generation and model training
Jetson Orin and IGXRuggedized edge inference at the station
TensorRT and DeepStreamMulti-camera pipelines at line speed
HoloscanLow-latency camera and sensor I/O
NeMo, NIM and TritonThe Pilora copilot, served on-premises or in the cloud
OmniverseLine and station digital twins in Twinix
NVIDIA AI EnterpriseSupported deployment for plant IT and OT
  • Synthetic runs, training jobs and deployments as resources
  • Webhooks for verdicts, drift and retrain events
  • Planned: Python SDK and OPC UA companion spec
POST /v1/synvex/runs
# Start a synthetic run for one SKU, then train an adapter on it
curl -X POST "https://api.viseena.com/v1/synvex/runs" \
  -H "Authorization: Bearer $VISEENA_TOKEN" \
  -d '{
    "sku": "AX-204",
    "cad": "cad/ax-204-rev-c.step",
    "material": "anodized-al-6061",
    "defects": ["scratch", "dent", "burr", "contamination"],
    "variants_per_class": 2000,
    "then": { "coreon_train": true }
  }'

# Response (abridged, example values)
{ "run": "syn_18", "status": "queued", "frames_planned": 8000 }
Questions

Common questions

Can a model trained on synthetic defects work on a real line?

That gap between simulation and reality is the main technical risk, and we treat it as one. Coreon aligns synthetic training to a small set of real frames, and every pilot starts by measuring accuracy on real parts before any decision is automated.

Do we need to replace our cameras?

Not necessarily. Veridex works with industrial GigE Vision and USB3 Vision cameras. Twinix can tell you whether your current placement and lighting can see the defects you care about.

Does inspection data leave the plant?

No by default. Veridex runs on the edge node beside the line. You choose whether any frames are sent to training, and where that training runs.

Are you live with customers?

No. Viseena is pre-launch. We are recruiting a small number of design partners for 2027 pilots.

Next step

Show us one hard-to-catch defect

Send images of the part and the defect. We will tell you whether a twelve-week pilot on your line can catch it.