AI Computer Vision Engineer (Mid-Senior)

EMPLOYMENT TYPE: Contract - Initial 12 months with possibility of extension based on delivery COMPANY: Vito Solutions CLIENT: An innovative technology company operating at the intersection of data science, biological research, and large-scale production systems. WORKING MODEL: Fully remote JOB OVERVIEW: Reports to: Lead Solutions Architect The Mission: Build and ship the computer vision models that turn poultry processing line video into real-time decisions: carcass defect detection, contamination flagging, operator cycle times, yield, and product tracking. Your models will run against a live line, not a static dataset. DESCRIPTION OF POSITION: What you own: Detection, segmentation, and tracking models for carcasses, defects, operators, and product on the line. Model accuracy against defined thresholds per use case: fecal and foreign-material detection, carcass defects, deboning quality, shackle occupancy, and operator productivity. Real-time performance. Models must keep up with line speed without dropping frames, whether we deploy on edge devices or a central GPU server. The full loop: data pipeline, training, evaluation, drift monitoring, retraining triggers. Partnering with the annotation lead to define labeling guidelines and mine hard edge cases, because the poultry dataset is the moat and it doesn't exist off the shelf. What you ship in the first 90 days: Days 1 to 30: reproduce a baseline detection model on the client's poultry dataset and get it running at target frame rate on our chosen inference target. Days 31 to 60: hit accuracy threshold on the first production use case, defect or contamination, validated against SME-labeled ground truth. Days 61 to 90: deploy that model to a pilot rig or live line- and stand-up drift monitoring so we know when it degrades. KNOWLEDGE AND SKILLS: Stack you'll work in: PyTorch and standard CV architectures (YOLO, RT-DETR class models) Annotation pipeline (CVAT or Roboflow or equivalent) Docker Supabase and Next.js control plane for surfacing results Inference hardware and runtime are still being decided and helping us make that call is part of the role. How we measure you: Model accuracy in production against threshold, inference latency at line speed, and time from new use case to deployed model. Not lines of code, not papers read. EXPERIENCE REQUIRED: 3 to 6 years experience building computer vision in production, not just research or coursework. Shipped real-time inference on video, not just batch image classification. Strong PyTorch. Fluent with detection and segmentation architectures (YOLO family, RT-DETR, Mask R-CNN or equivalent) and multi-object tracking (ByteTrack, BoT-SORT or equivalent). Model optimization for real-time and resource-constrained targets: quantization, pruning, and runtime tuning. Hardware-agnostic thinking matters here because we have not locked our deployment platform. Comfortable owning the full loop: data, training, eval, deployment, monitoring. Rigorous about evaluation. You know why mAP alone lies and you design metrics that match the business outcome. ADVANTAGEOUS SKILLS: Edge inference experience on any platform (NVIDIA Jetson, Hailo, Coral, or GPU servers) and the judgment to help us choose the right one. Runtime and pipeline frameworks such as TensorRT, ONNX Runtime, OpenVINO, or DeepStream. Industrial, manufacturing, food processing, or agriculture vision deployments. Vision-language models for scene-level reasoning or root-cause narration. Active learning, synthetic data, or transfer-learning toolkits. **Please note: If you have not heard from us within 2 weeks, please consider your application unsuccessful.