Mission
Translate STEP/IGES files into tolerance-aware, machine-interpretable geometry that powers the entire D2R alignment and QC pipeline. Your differentiation is native 3D CAD matching – if this layer is weak, the platform becomes another generic vision tool.
System Ownership
- Primary: CAD parsing engine (STEP, IGES, IFC, DWG → internal geometry representation)
- Primary: Feature extraction pipeline (holes, edges, weld seams, surfaces, datums)
- Primary: GD&T tolerance mapping (geometric dimensioning and tolerancing interpretation)
- Primary: Automated geometry comparison / CAD diff engine
- Secondary interface: CV team (they consume your mesh/point-cloud representations for alignment)
- Secondary interface: MR team (they render your extracted geometry as overlays)
- Does NOT own: Point cloud registration (CV team), MR rendering (MR team), deviation analytics (AI team)
What You Will Build
- CAD parsing engine – Ingest industrial CAD files (STEP, IGES, IFC, DWG) from major authoring tools (SolidWorks, AutoCAD, Revit, Inventor, Creo). Extract B-Rep geometry, assembly structure, and metadata. Handle files from 1MB to 2GB.
- Feature extraction system – Automatically identify and classify geometric features: holes (through, blind, counterbore), edges (fillet, chamfer, sharp), weld seams, surfaces (planar, cylindrical, freeform), and datum references.
- GD&T tolerance mapping – Parse GD&T annotations from CAD files (PMI data). Map tolerances to extracted features. Build a machine-readable tolerance model that the deviation detection system can evaluate against.
- Automated CAD diff engine – Compare two revisions of the same CAD model. Identify what changed: added features, removed features, dimensional changes, tolerance changes. Output a structured diff report.
- Geometry export pipeline – Convert internal representation to formats consumable by downstream systems: tessellated mesh for MR rendering, oriented point cloud for CV alignment, feature list for QC logic.
- Scalable mesh simplification – LOD generation for large assemblies. Full-detail for alignment, simplified for MR rendering, ultra-simplified for thumbnail previews. Maintain feature fidelity at each level.
Core Technical Responsibilities
- Build and maintain the CAD ingestion pipeline using Open Cascade Technology (OCCT) or equivalent geometric kernel
- Implement B-Rep traversal: iterate over solids, shells, faces, wires, edges, vertices. Extract topology and geometry at each level
- Build the feature recognition system: pattern-match geometric features from B-Rep topology (e.g., a cylindrical face bounded by planar faces = a hole)
- Parse PMI (Product Manufacturing Information) and GD&T annotations from STEP AP242 files – not all CAD files have clean PMI, build fallback heuristics
- Implement tessellation with controllable chord error for mesh generation – balance mesh density against downstream performance requirements
- Build the CAD diff algorithm: geometric comparison using shape signatures, topological comparison using B-Rep graph matching
- Handle degenerate geometry: zero-thickness faces, self-intersecting surfaces, invalid trim curves – CAD files from the field are not clean
Required Technical Mastery
- Computational geometry: B-Rep (Boundary Representation), NURBS surfaces, constructive solid geometry (CSG), Voronoi diagrams, Delaunay triangulation
- Geometric kernels: Open Cascade Technology (OCCT), CGAL, or equivalent. Deep API knowledge, not wrapper-level usage
- CAD formats: STEP (AP203, AP214, AP242), IGES, IFC (for construction), DWG/DXF. Understanding of format limitations and data loss during conversion
- GD&T: ASME Y14.5 / ISO 1101 tolerance specification. Datum reference frames, feature control frames, MMC/LMC modifiers
- Mechanical engineering fundamentals: Ability to read and interpret engineering drawings. Understanding of manufacturing processes (machining, welding, casting) that create the features you're extracting
- Languages: C++ (primary – OCCT is C++ native), Python (scripting, pipeline orchestration, prototyping)
- Mesh processing: Tessellation algorithms, mesh simplification (quadric edge collapse, vertex clustering), mesh repair (hole filling, manifold correction)
- Linear algebra: Surface fitting, least-squares approximation, principal curvature computation
Production Challenges You'll Solve
- Dirty CAD files – A customer uploads a STEP file exported from SolidWorks 2018 via an intermediate IGES conversion. The file has 47 invalid trim curves, 12 self-intersecting faces, and 3 zero-thickness features. Your parser must not crash. It must extract usable geometry and flag issues.
- Missing GD&T – 60% of real-world CAD files have NO GD&T annotations embedded. The tolerance information exists only in a separate 2D drawing PDF. Build heuristic tolerance inference from geometry (e.g., a 10mm hole with H7 fit implies ±0.015mm tolerance).
- Assembly scale – A full factory layout CAD file: 50,000 parts, 2GB file, 200M faces. Your parser must handle this without loading the entire model into memory. Stream-process, extract relevant sub-assemblies, generate LODs.
- Version diff across authoring tools – A customer sends revision A from SolidWorks and revision B from Inventor. The internal representations differ. Your diff engine must compare geometry, not file format, and produce a meaningful change report.
- Feature ambiguity – Is a small cylindrical protrusion a dowel pin locator, a weld stud, or a cosmetic boss? Context matters. Build feature classification that uses surrounding geometry and assembly context to disambiguate.
Success KPIs
| KPI | Target | Measurement |
|---|
| CAD format support | STEP, IGES, IFC, DWG | Successful parse rate ≥ 95% on customer file corpus |
| Feature extraction precision | ≥ 90% correct | Validated against manually annotated test CAD files |
| GD&T parse accuracy | ≥ 85% for AP242 files with PMI | Compared against manual GD&T interpretation |
| Parse latency | < 30s for files ≤ 500MB | End-to-end: file load → geometry export |
| Diff accuracy | ≥ 90% changed features detected | Compared against known revision pairs |
| Crash rate on malformed files | 0% | Parser must gracefully handle all inputs |
Failure If Underperforming
- CAD parser crashes on malformed files → system appears fragile. Customers with legacy CAD workflows (most industrial customers) cannot use the platform.
- Feature extraction misses critical features (holes, datums) → GD&T tolerance mapping is incomplete → QC pass/fail decisions are unreliable.
- No GD&T fallback → 60% of customer files produce no tolerance data → the system can only do basic dimensional comparison, not true QC.
- The platform without strong CAD understanding becomes indistinguishable from generic point cloud comparison tools – the core competitive advantage evaporates.
Collaboration Interfaces
| With | Interface |
|---|
| Lead CV Engineer | You provide tessellated meshes and oriented point clouds for registration. Coordinate system and scale contract must be exact. |
| MR Systems Engineer | You provide LOD meshes for overlay rendering. Define the mesh complexity budget jointly. |
| Applied AI Engineer | Your extracted features and GD&T tolerances define the pass/fail criteria their anomaly models evaluate against. |
| Backend Engineer | CAD files are stored and versioned in their system. Your parser runs as a processing pipeline on upload. |
Why This Role Is Mission-Critical
Our differentiator is native 3D CAD matching. Not image comparison. Not 2D overlay. Actual geometric understanding: parse the CAD, extract the features, map the tolerances, and compare against reality at the geometry level. This is what separates D2R from every other AR inspection tool on the market. If this layer is shallow (just mesh rendering without feature understanding), the product loses its moat and becomes a commodity.
About Us
Building the D2R (Design-to-Reality) platform – sub-millimetre CAD alignment + edge AI + mixed-reality overlay for industrial field workers. Venture-backed, seed-stage, < 20 engineers.
- Location: Bangalore / Hyderabad
- Stage: Seed / Pre-Series A (venture-backed)
- Industries: Construction, Manufacturing, Infrastructure, Energy