Coreon · training and models
One defect model, adapted to each SKU in hours
Coreon fine-tunes a lightweight adapter for each part on top of a shared defect model, aligns it with a small set of real frames, and versions every model it ships.
Specification
- Training
- NVIDIA TAO Toolkit on DGX/HGX or cloud GPUs; DALI data loading
- Tasks
- Detection, segmentation, classification, anomaly and novelty
- Inputs
- Synvex datasets; real good and known-bad frames
- Outputs
- TensorRT-ready adapters for Veridex, with an evaluation report
What Coreon does
- 01
Shared defect backbone
Each SKU starts from a model that has seen many parts and defect types (foundation model planned).
Planned - 02
Per-SKU adapters
Lightweight adapters fine-tuned with NVIDIA TAO, so a new part takes hours rather than a fresh training run.
Core - 03
Sim-to-real adaptation
Style and lighting alignment closes the gap between rendered and real frames.
Core - 04
Detection to anomaly
Detection, segmentation, classification and an anomaly head for defects nobody has named yet.
Core - 05
Versions and rollback
Every model is versioned per SKU and station, with evaluation attached and one-step rollback.
Core - 06
Learns from uncertainty
Uncertain real frames are queued for review and folded into the next adapter.
Core
What Coreon produces
AX-204 evaluation on real frames
- Base model
- Defect backbone (planned)
- Adapter
- AX-204 v7
- Training data
- Synvex run 18 + 120 real
- Heads
- Detect, segment, anomaly
- Registry
- Staged for Line A
Inside Coreon
Pretrain
The backbone learns from synthetic and consented real data.
Adapt
TAO fine-tunes a per-SKU adapter on Synvex data.
Align
A few real frames close the sim-to-real gap.
Evaluate
The harness reports per-class accuracy on real frames.
Specification
Items marked planned are in design with pilot partners.
| Item | Detail |
|---|---|
| Training | NVIDIA TAO Toolkit on DGX/HGX or cloud GPUs; DALI data loading |
| Tasks | Detection, segmentation, classification, anomaly and novelty |
| Inputs | Synvex datasets; real good and known-bad frames |
| Outputs | TensorRT-ready adapters for Veridex, with an evaluation report |
| Registry | Per-SKU and per-station versions, rollback, approvals |
| Foundation model | Cross-customer pretraining with consent (planned) |
Coreon in the Viseena loop
- P01 · GenerateSynvex
Renders labeled defect images from CAD, materials and a defect taxonomy.
- P02 · TrainCoreon
Fine-tunes a shared defect model into a per-SKU adapter, closing the sim-to-real gap.
- P03 · InspectVeridex
Runs the model on edge hardware at the station, at line speed, on every part.
- P04 · ExplainPilora
Turns flagged frames into defect reports, triage and probable root cause.
- P05 · CommissionTwinix
Simulates cameras and lighting in a line twin before anything is mounted.
Uncertain and confirmed real frames from Veridex flow back to Synvex and Coreon. Pilora triggers retraining when it sees drift, and Twinix sends new scenarios to Synvex before a changeover.
About Coreon
Is our data used to train other customers' models?
Not without written consent. Your images and adapters stay yours. The shared backbone only learns from data customers opt in to share.
What accuracy should we expect?
We will not quote numbers before pilots. Each pilot reports per-class accuracy measured on your real parts.
Can we export the model?
Adapters are exported as TensorRT engines for Veridex. Other export formats can be discussed for pilots.
Try Coreon on one SKU
Pilots start with one part, one defect list and accuracy measured on real frames.