Manual or destructive checks
Slow feedback, operator variability, and sampled evidence can leave the majority of production unseen.
AI-native inline CT inspection
See every part. Decide before it leaves the line.
ScanLine is building a physics-informed AI and X-ray CT system for fast, trustworthy 3D quality decisions—designed to move inspection from an offline sample to an inline control loop.
Physics-informed reconstruction
Uncertainty-aware decisions
Hybrid edge + cloud path
Concept visualization
Sparse acquisition
Measure only what the decision needs
2–5 s target
For a defined part and defect class
Low confidence?
Rescan or route to human review
The inspection gap
Factories still trade depth for speed.
The defects that matter are often internal and three-dimensional, while the tools that see them are commonly too slow or too remote from the production decision.
Offline industrial CT
Rich 3D evidence arrives too late when scan, reconstruction, and interpretation take longer than the takt time.
Fast 2D inspection
High throughput can come at the cost of depth ambiguity, occlusion, and incomplete localization of internal defects.
Our thesis: co-design acquisition, reconstruction, and the quality
decision—then optimize the whole loop for the defect that must be
caught.
The ScanLine loop
From X-ray projections to an auditable action.
AI is not a cosmetic add-on. It connects sparse measurements, physical consistency, defect evidence, uncertainty, and factory feedback in one decision system.
Part + defect truth
Define the decision, minimum defect, and cost of each error.
Sparse acquisition
Select geometry and projections around the target decision.
Physics-informed AI
Reconstruct or infer while enforcing projection consistency.
Defect decision
Return localization, confidence, and a traceable pass/fail basis.
Factory feedback
Rescan, review, and feed evidence into MES/QMS and process control.
Cycle time
Minimum detectable defect
Recall
False-positive rate
Dose
Uncertainty calibration
Evidence, separated from ambition
Built by an operator who has taken inline CT to the line.
The founder’s prior system experience is established. ScanLine’s new product, code, datasets, and IP path are being built independently—and the production-speed performance bridge remains a testable milestone, not a claimed result.
End-to-end inline CT experience
Zhihua previously led AI back-end, reconstruction, and system integration work for an inline CT platform, including line-side deployment.
An independent ScanLine stack
New acquisition design, reconstruction and decision software, data contracts, and a clean IP route are under development.
Production-speed acceptance
The 2–5 second cycle target and defect sensitivity must be validated jointly on a defined part, defect, and line.
Trustworthy by design
A quality decision needs more than a sharp image.
Projection consistency
Check AI output against the measurements and the physics.
Defect-grounded truth
Validate against metrology, destructive sectioning, or a qualified reference method.
Uncertainty and drift
Detect out-of-distribution parts and surface low-confidence decisions.
Human-safe fallback
Escalate ambiguous cases to rescan or expert review instead of hiding uncertainty.
Two routes, one intelligence layer
Start integrated. Remain hardware-flexible.
The strongest route is a co-designed inspection system. An independent software route can upgrade compatible CT hardware and avoids making every deployment depend on a new scanner.
Champion route
Integrated inline CT system
Co-design source, detector, motion, sparse acquisition, AI reconstruction, and the final quality decision around one production problem.
Independent alternative
AI software for compatible CT
Deliver reconstruction, defect decision, uncertainty, and workflow integration as a hardware-compatible intelligence layer.
The decisive first wedge
One part. One defect. One line.
A narrow first deployment makes the hardest claim falsifiable: can sparse CT and physics-informed AI beat a dense baseline on cycle time while meeting the same defect decision threshold?
1 part
A stable geometry with a real production bottleneck
1 defect
A measurable internal failure mode with qualified truth
1 line
A takt-time target and an accountable process owner
Lock the part, defect taxonomy, truth method, dense baseline, and acceptance metrics.
Acquire sparse/dense pairs, train the physics-informed pipeline, and calibrate uncertainty.
Run a blind comparison and choose: deploy, repair the observable, or stop that wedge.
Candidate applications
Begin where hidden defects carry a high cost.
These are target application classes—not customer or deployment claims. The first commercial wedge will be selected by data access, defect truth, takt time, and economic value.
Battery cells
Internal alignment, folds, contamination, and structural anomalies.
Castings and welds
Porosity, cracks, inclusions, and internal joining defects.
Electronics
Hidden interconnect, packaging, and assembly anomalies.
Two-person core team
CT systems experience meets medical imaging AI.
A compact technical team spanning inline CT architecture, reconstruction, computer vision, and medical image analysis.
Zhihua Liang
Founder · CT Systems & AI
Previously led AI back-end, reconstruction, and system integration for the DeltaRAY X100 inline CT platform. Focused on turning industrial imaging into a production decision system.
LinkedIn ↗Dr. Juan Liu
Core Team · AI Medical Imaging
PhD researcher in AI medical imaging, contributing image analysis, model development, experimental design, and clinical imaging perspective to the ScanLine technology path.
Inline CT architecture
Reconstruction
Physics-informed learning
Medical image analysis
System integration
Execution roadmap
Turn the open bridge into measured evidence.
The current priority is not a broad product catalogue. It is one defensible production result that becomes a repeatable system.
Architecture + wedge
Independent product architecture, pilot definition, and data/IP contracts.
Decisive benchmark
Dense-vs-sparse blind comparison on one part and one defect.
Line-side prototype
An integrated prototype tested in a real production environment.
Repeatable deployment
A qualified benchmark, reusable workflow, and first commercial deployments.
AI and cloud roadmap
Build a learning inspection system, not an isolated model.
We are evaluating the right Google Cloud and Gemini stack for scalable 3D training, multimodal manufacturing knowledge, and hybrid deployment where production data may need to stay on site.
Multimodal quality copilot
Connect volumes, projections, defect evidence, process logs, and engineering knowledge.
Scalable 3D model development
Evaluate Vertex AI and accelerators for training, evaluation, and reproducible model governance.
Hybrid edge + cloud operations
Keep real-time inference near the line while aggregating permitted evidence for fleet learning.
Build the first proof with us
Have one costly hidden defect?
We are looking for a design partner who can define one part, one defect class, a qualified truth method, and a real takt-time constraint.
90-day scoped technical pilot
Joint acceptance metrics agreed before data collection
Clear data, publication, and IP boundaries