Factory AI data, labeled by engineers.
Industrial models fail on the defect nobody photographed and the part nobody modeled. We source the data you lack and have engineers label it to your tolerances.
Manufacturing, with proof.
Manufacturers use models for visual inspection, design assistance, process documentation, maintenance support and robotic assembly. The data problem is consistent across them. Defects are rare by design, so a good line produces few examples of the faults you most need to catch. CAD models and process know-how live in proprietary systems. HUMXN sources defect imagery, 3D and CAD data, technical documents and operator demonstrations from consenting creators and commissioned experts, then has engineers label them against your specification. A pilot batch comes first.
Engineering data needs engineering review. Our mechanical and manufacturing specialists label defect type, location and severity, annotate CAD features and tolerances, and write the reasoning behind design and process decisions. Labels record their source, and a second expert reviews disputed items, which matters when the line between a cosmetic mark and a reject is a judgment call. 3D files in STL, OBJ, PLY, GLB, GLTF, DXF and point-cloud formats are shape-fingerprinted at intake, so duplicate parts do not inflate your dataset.
Design files carry real intellectual property. Each catalogue work records its AI-training permission and licence terms, and exclusive licences keep commissioned data yours. Documents and tables are scanned for personal data and masked before delivery. Every item has an Ed25519-signed provenance record and appears on a signed receipt by ID and hash. For teams whose models will sit inside quality systems, that gives a documented chain from source to training set that supports your internal validation work.
Built to your specification.
Every engagement starts from a written spec and a pilot batch. These are the most common requests we source for manufacturing.
Defect inspection imagery
Photographs of surface, assembly and weld defects across materials, labeled for type, location and severity by engineers.
CAD and 3D part data
Parts and assemblies in STL, OBJ, PLY, GLB, GLTF and DXF, with feature and tolerance annotation for design models.
Technical document labeling
Work instructions, maintenance manuals and specifications structured and labeled, with personal data scanned and masked.
Assembly teleoperation demonstrations
Robot demonstrations of insertion, fastening and kitting recorded by trained operators, with success and failure labels.
Engineering reasoning traces
Expert-written explanations of design choices, root-cause analysis and process troubleshooting, for engineering assistants.
Maintenance and diagnostic data
Expert-labeled audio, video and sensor recordings of equipment conditions for predictive maintenance and diagnostic models.
01Rare defects
Good lines produce few faults. Sourcing defect examples to a specification gives inspection models the cases they must catch.
02Judgment at the margin
Cosmetic or reject is often a judgment call. Engineer labels and second-expert review keep that line consistent.
03Design IP
Licence terms and training permission are bound to each file, and exclusive licences keep commissioned data out of other hands.
Manufacturing: asked often.
Can you provide images of specific defect types?
We source against a written defect specification covering materials, defect classes, lighting and camera setup. Images come from consenting creators and commissioned work, and engineers label type, location and severity. Where a defect class is genuinely hard to source, we say so during scoping. The pilot batch shows what coverage looks like before you scale.
Which CAD and 3D formats do you support?
STL, OBJ, PLY, GLB, GLTF, point clouds and DXF. Each file is shape-fingerprinted at intake so duplicate and near-duplicate parts surface early. Feature, tolerance and assembly annotations are delivered alongside in JSON Lines or Parquet, with Croissant 1.0 metadata describing the dataset if your catalog uses it.
How do you protect design intellectual property?
Every catalogue work carries its AI-training permission and licence terms, so you only receive files cleared for your use. Commissioned data can be delivered under an exclusive licence. Each item is listed on a signed receipt by ID and hash, and confidentiality terms for your project are set out in the agreement.
Who labels manufacturing data?
Credential-checked mechanical, quality and manufacturing engineers, CAD designers and maintenance technicians, matched to the task. Labels record whether they came from a creator, an AI suggestion or expert review. Borderline items, such as cosmetic versus functional defects, go to a second expert under the criteria in your spec.
How is manufacturing data priced?
Price depends on how hard the data is to source, annotation density, the expertise required, review depth and licence. Rare defect classes and detailed CAD annotation cost more than common images with simple labels. Volume and standing orders reduce the unit price. We quote once the spec is agreed.
Tell us what your model needs to learn.
Send a brief in five minutes. A data lead replies within one business day with questions and a first sourcing plan.