Machine Learning Engineer for Signal Training

The deliverable of this role is a training method, not a trained model. We are not asking you to maximise the accuracy of a classifier. We are asking you to implement a specific, non-standard way of training a model on 1-D physiological signals, and to prove rigorously that the training method itself is what produces the effect. The trained model is only the vehicle used to demonstrate and measure the method; it is not the output. The reusable, documented training method is. What you’ll build - Baselines (for comparison only): standard model training and standard data augmentation for a 1-D signal detection task. These exist purely as the thing our method must beat. - The custom training method (the core deliverable): implement a paired / factorial, intervention-based training scheme that we specify on start. This is custom loss-function and training-loop work, not configuration of an off-the-shelf model. - Ablations and negative controls: remove each component of the method, and run negative controls (e.g. scrambled pairings, shuffled-label probe), to show the effect comes from the method and not from extra data, extra parameters, or leakage. - Reporting: metrics plus a short technical write-up documenting the method and what the experiments show. What is explicitly NOT the deliverable - A highly tuned or production-ready model. - A high benchmark-accuracy score. - Clinical validation (handled in-house). What you must have - Strong Python and deep learning (PyTorch or TensorFlow). - Demonstrable experience writing custom loss functions and training loops — not only fine-tuning pre-built models. This is the essential skill. - Experience with time-series / 1-D signals (ECG or physiological signals a plus). - Rigorous experimental practice: patient-level splits, leakage control, ablations, reproducibility (fixed seeds, re-runnable scripts). Deliverables - A reproducible codebase implementing the training method and the baselines. - An ablation and negative-control results table showing the method is responsible for the effect. - A short methods-and-results report (2–3 pages) that documents the training method as the asset. Data: will be supplied, along with the exact training specification, on start under NDA.