Senior Computer Vision & Mobile AI Engineer - On-Device Pose Estimation for Workplace Safety

We're building a workplace-safety system that detects people and tracks their body posture in real time to flag unsafe movements - bad lifting form, dangerous bends, awkward positions that lead to injury. It has to run on-device (edge/mobile), not on a server. Before we commit to the full build, we're bringing in a senior engineer to lay the technical foundation and prove it's feasible. Milestone 1 (this contract) - technical foundation + mini pilot: Choose the right on-device pose-estimation approach for our case (YOLOv8-pose / MoveNet / MediaPipe / other) and justify the tradeoffs for edge deployment. Define which keypoints and body angles actually matter for safety logic (spine bend, lifting posture, joint angles) - what we need to track and why. Assess real-time feasibility on a mid-range phone - realistic framerate, model size, accuracy limits. Deliver a mini proof: run your chosen pose model on a short sample video we provide and show the skeleton tracking working. Flag the real risks before we build - where accuracy breaks, where latency bites, what will be hard on-device. For someone who's shipped pose estimation to mobile, this is a focused piece of work with clear written recommendations plus a working mini-demo. Requirements: Real experience with pose estimation / keypoint detection in production, not just object detection. Hands-on on-device deployment (CoreML, TFLite, ONNX Runtime Mobile) with the quantization/latency tradeoffs that come with it. Able to look at a real-world CV problem and predict where it breaks before writing the full system. If Milestone 1 goes well, this leads directly into the full build - real-time skeleton tracking from the camera, safety-rule logic (unsafe pose → alert), and on-device optimization. We're looking for one specialist to own this long-term. Please share a pose-estimation or on-device computer vision project you've shipped — and one technical tradeoff you'd flag for a real-time safety use case.