AI Engineer - Computer Vision

Work Description :

• Train and fine-tune instance segmentation and object detection models (Mask R-CNN, YOLO-family) to detect irrigation and landscape symbols across mixed symbol libraries, varying scales and rotations.
• Build the hatch and bed-line segmentation pipeline that measures areas for sod, seed, mulch, paving and similar, including cases where identical hatches carry different labels or appear nowhere in the legend.
• Design the tiling and preprocessing strategy for large-format plan sheets, and own the trade-off between tile count, detection recall and inference cost.
• Implement calibrated confidence scoring, and the "unknown - needs review" path so unfamiliar symbols surface instead of being misclassified with false confidence.
• Run the three fine-tuning rounds : train, evaluate against acceptance criteria, error-analyse which classes underperform, and carry a written remediation plan into the next round.
• Deploy models to Amazon SageMaker asynchronous endpoints with scale-to-zero, and keep releases immutable and version-tagged so any release can be rolled back.
• Work with the QA engineer on per-class acceptance measurement, no-regression checks and the calibration and held-out split.

Skills required :

• Strong Python, and production experience with PyTorch on detection or segmentation tasks - not just notebook experiments.
• Hands-on work with Mask R-CNN, YOLO, Detectron2 or similar, including training on custom datasets and diagnosing class-level failure.
• OpenCV and image preprocessing : tiling, rotation, scale normalisation, morphology.
• PDF and raster handling at scale (PyMuPDF, pdf2image, Poppler, Pillow).
• A working understanding of detection metrics - mAP, IoU, precision and recall per class - and why aggregate accuracy hides per-class regression.
• AWS : S3, SQS, and either SageMaker or comparable managed inference.

Nice to have :

• Experience reading engineering, architectural or CAD-derived drawings.
• Active learning or human-in-the-loop correction pipelines.
• Model optimisation for inference: TensorRT, ONNX, quantisation.
• Annotation tooling and labelling workflow design.