Our health apps guide patients through rehabilitation exercises, watch them train via camera and tell them what to correct. Today that runs on biomechanical rules written by hand for each exercise. Your mission: Replacing them with a model that evaluates movement quality directly from keypoints, handling the following challenges:
- Scale: Hundreds of exercises, several fault classes each, and what counts as a fault shift with the impairment. Adding new exercises or error classes should stay cost efficient – with one model per exercise we don’t get there, so generalization is needed
- Constraints: Edge deployment directly on mid-range Android or iOS devices. You will handle noisy keypoints, occlusion, dropped frames, and low-latency edge processing to give patients instant feedback to their exercise repetition.
- Reference: Correct execution is a clinical judgement, and our physiotherapists and doctors are here to define it with you. Turning that judgement into something a model can learn from, and into an evaluation we can trust, is part of the research question
Having end to end ownership for the feedback model also includes:
- Identify, extract, and process data from relevant sources, and shape the annotation strategy
- Research new approaches and turn them into solutions
- Design and implement deep learning pipelines with a focus on computer vision, from data through training and evaluation to deployment.
- Create tests and evaluations that measure model, but also production performance
With a data science team of three, you will also do hands-on work across the team, especially supporting the optimization of our body and hand tracking models.
