Hoa N.

If your model isn’t performing well, the problem is often the data — I help fix that. I specialize in data-centric computer vision systems: improving detection accuracy, reducing false positives, refining datasets, and deploying stable real-time AI pipelines on edge and mobile devices. I build end-to-end computer vision workflows from dataset preparation and model training to real-time Android deployment. What I help with: ✓ Reducing false positives and missed detections ✓ Dataset QA, cleaning, validation, and deduplication ✓ Improving label consistency across large-scale datasets ✓ Building feedback loops between model predictions and dataset correction ✓ Embedding-based similarity and clustering for duplicate detection ✓ Segmentation mask processing and structured object extraction ✓ Real-time object detection and tracking systems ✓ Improving tracking stability and frame-to-frame consistency ✓ Edge/mobile AI inference optimization ✓ Real-time Android deployment workflows Real-world experience: ✓ Built and deployed computer vision systems for fitness applications ✓ End-to-end pipeline development: dataset preparation → training → inference → Android deployment ✓ Real-time on-device inference pipelines ✓ Barbell tracking and repetition counting ✓ Skeleton-based motion analysis ✓ Equipment classification and tracking consistency ✓ Turning raw detections into stable, usable systems Technical stack: ✓ YOLO (training, fine-tuning, evaluation) ✓ OpenCV, PyTorch, Ultralytics YOLO, SAM ✓ TensorFlow Lite (TFLite) and ONNX deployment workflows ✓ Android Studio, CameraX ✓ CVAT, Label Studio, Roboflow, Labelbox, Supervisely ✓ QGIS, GeoTIFF, GeoJSON, MultiPolygon