Mission
Own the core measurement engine that converts LiDAR + camera input into sub-millimetre CAD-aligned dimensional intelligence. You are responsible for the accuracy layer that the entire D2R platform stands on – if alignment drifts by 3mm, the product's value collapses.
System Ownership
- Primary: 3D perception pipeline (point cloud acquisition → registration → alignment → deviation output)
- Primary: CAD-to-scan alignment engine (ICP variants, global registration, RMS scoring)
- Primary: Sensor fusion layer (LiDAR + RGB-D + IMU synchronisation and calibration)
- Secondary interface: Edge AI team (your alignment outputs feed their inference models)
- Secondary interface: MR team (your deviation maps drive their overlay rendering)
- Does NOT own: On-device model deployment (Edge AI), MR rendering pipeline, cloud analytics
What You Will Build
- 3D registration engine – ICP-based point-to-point and point-to-plane alignment between CAD geometry and real-world LiDAR scans. Handle symmetry ambiguity in repeated structural elements (beams, pipes, flanges).
- Point cloud processing pipeline – Multi-resolution processing for scan data exceeding 1M points per cycle. Voxel downsampling, statistical outlier removal, normal estimation at scale.
- Depth calibration & sensor fusion – Temporal and spatial synchronisation of LiDAR, RGB-D, and IMU streams. Compensate for sensor desync that causes ghost points.
- Deviation detection & tolerance mapping – Compute per-point and per-feature deviations against CAD reference. Generate deviation heatmaps with GD&T-aware tolerance bands.
- Large-scale object accuracy – Maintain sub-3mm alignment accuracy on structures exceeding 2 metres. Solve drift compensation for large-scale industrial objects.
- On-device inference optimisation – Collaborate with Edge AI to ensure perception models meet the 200ms latency budget on target hardware.
Core Technical Responsibilities
- Implement and optimise ICP registration variants (point-to-point, point-to-plane, generalised ICP) with convergence guarantees for industrial geometry
- Build RANSAC-based robust estimation for initial alignment under high outlier ratios (dust, occlusion, partial scans)
- Design the CAD-to-scan coordinate system alignment pipeline – handle scale mismatches between CAD units and scan units
- Implement occlusion handling: detect partially scanned regions, prevent incorrect convergence, flag low-confidence alignment zones
- Build multi-format CAD ingestion (STEP, IGES, IFC, DWG) → extract mesh/point-cloud representations for alignment
- Profile and optimise memory usage for point cloud operations on ARM-class edge hardware
- Establish automated regression testing for alignment accuracy across a growing library of industrial test cases
Required Technical Mastery
- 3D geometry fundamentals: Rigid body transformations (SE(3)), homogeneous coordinates, quaternion rotations, Rodrigues' formula
- Registration algorithms: ICP (all variants), RANSAC, Fast Global Registration, Super4PCS, feature-based registration (FPFH, SHOT descriptors)
- SLAM fundamentals: Visual-inertial odometry, loop closure, pose graph optimisation – you don't own SLAM, but you must understand drift to compensate for it
- Libraries: Open3D, PCL (Point Cloud Library), Eigen, OpenCV, PyTorch/PyTorch3D
- CAD formats: STEP, IGES parsing and mesh extraction. Working knowledge of Open Cascade or equivalent
- Languages: Python (prototyping + pipeline orchestration), C++ (performance-critical inner loops, memory-managed point cloud operations)
- Linear algebra: SVD for point set registration, least-squares optimisation, covariance estimation for alignment confidence scoring
- Edge deployment awareness: Model profiling, latency budgeting, understanding of TensorRT/ONNX export for perception models
Production Challenges You'll Solve
- Symmetry ambiguity – A factory floor has 200 identical I-beams. Your registration engine converges on the wrong one. Build disambiguation using contextual spatial constraints and sequential scan ordering.
- Partial occlusion – A welder is standing in front of the structure being scanned. 40% of the point cloud is missing. Your alignment must still converge within tolerance or explicitly flag insufficient data.
- Scale mismatch – The CAD file uses millimetres, the LiDAR outputs in metres, and the RGB-D camera uses yet another scale. Build robust unit detection and automatic normalisation.
- Dynamic lighting – Outdoor construction site, midday sun causing RGB-D depth holes. Welding flash on a shop floor saturating sensors. Your pipeline must degrade gracefully, not crash.
- Drift on large objects – Aligning a 10-metre pipe rack. Accumulated registration error exceeds 3mm by the far end. Implement hierarchical or segmented registration to bound cumulative drift.
Success KPIs
| KPI | Target | Measurement |
|---|
| Alignment RMS error | < 3mm | Measured against ground-truth metrology scans on 20+ test objects |
| Alignment latency | < 250ms | End-to-end on target edge hardware (per alignment cycle) |
| Point cloud throughput | ≥ 1M points/cycle | Processing + registration combined |
| Convergence rate | ≥ 95% | Percentage of scans that converge within 3 ICP iterations |
| False alignment rate | < 1% | Alignments that converge but on the wrong object/feature |
| Regression test pass rate | 100% | Automated test suite across all reference geometries |
Failure If Underperforming
- Alignment accuracy degrades below 3mm → product is unusable for precision manufacturing QC. Customers cannot trust deviation measurements.
- Alignment latency exceeds 250ms → MR overlay lags behind physical movement. Field operators stop using the system.
- Symmetry ambiguity unresolved → false alignments cause incorrect pass/fail QC decisions. One bad call on a safety-critical weld = liability.
- Partial occlusion crashes pipeline → system appears unreliable in real factory conditions. Customer churn.
Collaboration Interfaces
| With | Interface |
|---|
| Edge AI Engineer | You provide aligned point clouds + deviation maps. They run inference on your output. Joint latency budget: your 250ms + their 200ms. |
| MR Systems Engineer | Your alignment transform + deviation heatmap feeds their overlay renderer. Coordinate system handoff must be exact. |
| CAD Geometry Engineer | They own CAD parsing and GD&T extraction. You consume their mesh/point-cloud output for registration. |
| Applied AI Engineer | Your deviation data feeds their anomaly detection models. Data format contract must be stable. |
Why This Role Is Mission-Critical
Our entire value proposition is sub-millimetre CAD-to-reality alignment. This is not a feature – it is the product. Every downstream system (MR overlay, QC logic, analytics, compliance reporting) depends on the accuracy of what you build. If this layer is weak, the platform is an expensive camera app. If this layer is strong, it replaces manual metrology across construction, manufacturing, and energy – a market measured in billions.
You will be the technical authority on 3D perception at a company where 3D perception IS the core IP.
About Us
We're building the D2R (Design-to-Reality) platform – the first system combining sub-millimetre CAD alignment, on-device edge AI, and mixed-reality overlays for industrial field workers. Venture-backed, seed-stage, < 20 engineers. Your code ships to factory floors and construction sites, not dashboards nobody reads.
- Location: Bangalore / Hyderabad
- Stage: Seed / Pre-Series A (venture-backed)
- Industries: Construction, Manufacturing, Infrastructure, Energy