Build Private YOLO Training Pipeline for Outdoor IR Bullet-Hole Detection

We are building Smart Target Systems, an outdoor IR-camera bullet-impact detection system. Our current CPU/OpenCV/Cognex-style machine vision approach has limitations, and we want to move toward YOLO/deep learning or a hybrid YOLO + OpenCV approach. Important: our production target/bullet-shot dataset must remain private. The freelancer will not receive the full raw dataset. After NDA, we may share only a tiny sanitized sample for validation. Full training must be possible on our own machines by our internal engineers. We need a user-friendly local training pipeline that non-ML engineers can run repeatedly: import images/video frames, label or review bullet-hole annotations, export YOLO datasets, train Ultralytics/PyTorch YOLO models, evaluate false positives/negatives, inspect predictions, export weights/config, and package trained models. Please also include a small inference/deployment demo on sample or synthetic images/video so our team can understand how the trained model should work. Full production live-camera integration is not required in v1. Deliverables: - Local app and/or CLI for dataset import, labeling/review, training, evaluation, and export - Repeatable YOLO training workflow using Python, PyTorch/Ultralytics, and OpenCV where useful - Basic inference/deployment demo using sample or synthetic images/video - Hardware recommendations for training and deployment - Setup documentation and handover/training session for our engineers We are looking for someone with strong computer vision experience, especially small-object detection, YOLO model training, outdoor/IR imagery, false-positive reduction, and practical deployment handoff.