Machine Learning Engineer (3D Shape Classification & Assembly Intelligence) - FT - WW

Juliana Torrisi is in direct contact with the company and can answer any questions you may have. Email Juliana Torrisi, Recruiter

The work

We build software that turns 3D assemblies - meshes, CAD-derived geometry, and the messy reality of customer exports - into structured, queryable knowledge. Filenames lie. Export quality varies. Two parts that should be the same instance often arrive looking subtly different. Shape, treated rigorously, is the signal that holds up.

We are hiring an ML engineer to make that signal real: to turn derived geometric data into reliable classification, embeddings, and calibrated confidence that power deduplication, search, prioritization, and downstream reasoning. The goal is a product grounded in quantified 3D structure - not heuristics duct-taped to generic vision models.

What you will own

• The shape modeling roadmap, end to end. Problem framing, data, training, evaluation, production integration. You decide what gets built and what gets cut.
• Data and labeling for part and shape families. Taxonomies, weak labels from metadata and rules, clustering-assisted review, active learning, dataset versioning across heterogeneous customer geometry.
• Models that consume derived 3D data. Voxel occupancy, dimensional summaries, axis and shape descriptors, mesh-derived features - combined with render-based signals where they clearly help.
• Classifiers, embeddings, or hierarchical outputs with calibrated uncertainty where it matters.
• Production integration. Batch inference on large jobs. Clean schemas (predicted class, embedding vector, scores). Feature flags. Backward-compatible rollout.
• Honest measurement. Offline metrics (accuracy, calibration, long-tail behavior) and online product metrics (duplicate-handling quality, reduced ambiguity downstream, review throughput).
• Working contracts with adjacent systems. Mesh processing, rendering, analysis orchestration. You document the failure modes - bad meshes, degenerate solids, unit and scaling quirks - so they stay debuggable instead of becoming folklore., * A baseline shape model integrated in a controlled way, with agreed metrics moving the right direction.
• Documentation that downstream consumers of model outputs actually use - schemas, semantics, known limitations.
• A credible path to iterative improvement across new classes, customers, and failure modes - without the work collapsing back into one-off notebooks.

Why this role matters

Shape is the stable signal in a world of changing filenames and varying export quality. A real classification and embedding layer reduces redundant work, improves consistency across duplicate instances, and gives every downstream system - rule-based or model-based - grounded structure to build on. You will be central to making that layer load-bearing.

• Strong Python and production engineering habits. Tests, logging, reproducible runs - not as ceremony, as how you actually work.
• Solid ML fundamentals. Supervised learning, class imbalance, long tails, train/val/test discipline, and honest evaluation under domain shift.
• Practical fluency with 3D data. Volumes, meshes, or point clouds. You have dealt with normalization, alignment, sampling, and the numerical stability problems real geometry creates.
• At least one model shipped beyond a notebook. A batch job, a service, an embedded inference path - somewhere latency and resource use actually mattered.

Preferred

• 3D ML experience in Production. Voxel CNNs, point-cloud architectures, mesh and graph methods, modern - 3D representation learning. Familiarity with tessellation artifacts and what noisy real-world geometry actually looks like.
• MLOps fundamentals. Experiment tracking, artifact storage, reproducible environments, staged deployment.
• Industrial or mechanical domain exposure. Fasteners, brackets, housings, weldments - or a genuine appetite to ramp on the domain quickly with engineering partners.
• Interest in human-in-the-loop workflows. Review and correction loops that improve labels and models over time, not just dashboards that watch them.