Surface & cosmetic
Blemishes on an otherwise-good part, where the hard question is what counts as bad.
- Scratches, gouges, and tool marks
- Porosity, inclusions, and voids
- Contamination, residue, and staining
- Burn, discoloration, and finish variation
Machine Vision
Defect inspection that runs at line rate
We design, integrate, and tune in-line vision inspection on industry-standard hardware, typically Keyence and Cognex platforms, so defects are caught on the line instead of at final QC, a customer dock, or a warranty claim.
What in-line inspection actually catches
"Defect detection" covers several genuinely different problems, and they need different optics, lighting, and logic. Most lines need more than one of these, which is why the hardware choice follows the defect list rather than the other way around.
Dimensional & geometric
Measurement against nominal, where the tolerance band is the whole decision.
- Bore and outside diameters
- Flush and gap on mating surfaces
- Hole position and pattern spacing
- Flatness, warp, and coplanarity via 3D profiling
Assembly & presence
The part is fine, but something about the build is wrong.
- Missing or extra fasteners and components
- Wrong variant installed for the order
- Misorientation and flipped parts
- Seating, engagement, and weld presence
Marking & traceability
Verification that the part can be identified downstream and by the customer.
- Data matrix and barcode read and grade
- OCR of stamped or laser-marked characters
- Label presence, placement, and legibility
- Serial capture tied to the inspection record
Tolerance is a parameter, not a preference
Out of the box, most vision platforms give you one global sensitivity setting. That forces a bad compromise: tighten it and you scrap good parts, loosen it and you ship bad ones. The compromise gets worse the more variants share the line.
We build the control layer so the tolerance band is a first-class, per-feature parameter. A cosmetic scuff on a hidden face and a bore diameter on a safety-critical joint do not belong on the same threshold, and they should not have to share one.
- Independent bands per feature, so tightening one check never drags the others with it
- Per-variant and per-recipe values, so a shared line does not inherit one part's tolerances
- Real engineering units (mm, degrees, area, pixel count) rather than an opaque 1–100 sensitivity dial
- Adjustable without touching model code, so process engineers control the setting, not us
- Logged against every decision, so an audit can reconstruct which threshold was live when a part passed
There is a working demonstration of this on the home page: parts arrive with randomized surface spotting, get scanned, and are accepted or dropped through a reject hatch against a threshold you set with a slider, with yield recalculating live as you move it. That tradeoff between yield and escape rate is the real conversation on most lines.
Platforms we deploy on
In most engagements we are integrating and tuning proven hardware rather than building our own. That keeps you on a supported platform your maintenance team can get parts and help for, and it puts our effort where it changes the outcome: optics, lighting, recipe logic, tolerances, and the integration into your controls.
- Keyence IV series and VS line for area-scan inspection and 3D shape measurement
- Cognex In-Sight and comparable smart cameras for inspection and code reading
- Multi-camera and 3D profiling setups where one view cannot cover the feature set
- Edge inference where the decision needs to be local and deterministic
- Deep-learning inspection for defects that resist a rules-based description
We are vendor-agnostic by preference. The right camera is the one that resolves your smallest defect at your line speed and integrates with the controls you already run.
What a deployment looks like
1
Feasibility and sample study
We start with real parts, including the defective ones you have been saving. If the defect cannot be resolved reliably under controlled optics, no amount of software fixes it, and we would rather establish that in week one than after a purchase order.
2
Optics, lighting, and fixturing
The single largest determinant of a vision system's reliability. Lighting geometry, wavelength, and part presentation get selected against your actual defect list and line speed, not a datasheet.
3
Recipe and tolerance configuration
Inspection logic per part variant, with tolerance bands set per feature in engineering units and reviewed against your spec and your scrap history.
4
Controls and data integration
Pass/fail and measurement data delivered to the PLC, HMI, MES, or historian in the form your systems already expect, so results drive real actions like rejects, alarms, and routing.
5
Validation and handover
Run against known-good and known-bad samples, confirm the escape and false-reject rates against the agreed baseline, then hand over documentation along with the ability to adjust tolerances in normal operation.
Where it pays for itself
Inspection projects usually have a defensible number attached, which is why they often qualify for our profit-sharing model rather than a fixed fee:
- Scrap caught early instead of after value has been added downstream
- Containment and sorting labor avoided when a suspect lot would otherwise be inspected by hand
- Escapes prevented, where the real cost is the customer response, not the part
- Inspection labor reallocated from repetitive visual checks to work that needs judgment
- Throughput recovered when a manual inspection step is the constraint on the line
If the gain is measurable, we would rather share in it than bill for it. That is the whole idea behind our profit-sharing engagement model.
Have a defect you cannot catch reliably?
Tell us what the defect is, what it costs you, and what the line speed is. We will tell you honestly whether vision is the right answer.