Take an existing guardrail-defect-detection POC to production quality. First priority is cutting the model's false-positive rate (currently 83% on reviewed detections); then extend defect classification and condition ranking.
Must Have:
Computer Vision & Object Detection, Python (PyTorch/TensorFlow & OpenCV), YOLO/Detectron2 (Custom Model Fine-Tuning), Video Processing & Image Analysis, Model Evaluation & False-Positive Reduction, Production Codebase Development & Maintenance
Required:
4+ years hands-on computer vision / object detection in production (not research-only)
Strong Python, PyTorch or TensorFlow, OpenCV
Object detection frameworks (YOLO, Detectron2, or similar) incl. fine-tuning on custom data
Video processing experience (frame extraction, motion blur, variable lighting/weather)
Model evaluation / error analysis - precision-recall tuning, false-positive reduction
Comfortable working inside an existing codebase, not greenfield
Nice to Have:
AWS ML tooling (SageMaker, S3, Lambda)
Infrastructure/civil-inspection CV experience
GPS/GIS data handling or RTK GNSS/LiDAR exposure
Public sector client experience