ML Engineer (Movement Detection)

Kevych Solutions

Що робити

Increase accuracy and expand coverage to several hundred exercises. You will fine-tune existing models, redesign ensembling strategies, evaluate architectures suited for long-tail, imbalanced multi-class time-series classification, and adapt training regimes (data augmentation, sampling, curriculum learning) for newly annotated data.
Port the server-side inference pipeline to run live on iOS/Android devices under strict latency, memory, and compute constraints, handling real-world sensor noise and dropped samples through model export, quantization, and distillation.
Ensure all architectural and ensembling wins on offline accuracy respect mobile edge-budget constraints right from the design phase.

Hands-on experience around 5 years with PyTorch and PyTorch Lightning for building, training, and extending production pipelines, along with Python proficiency.
Strong applied deep learning background grounded in numerical/signal time-series data (IMU, EMG, audio, sensor fusion, biosignals, or industrial time-series) using architectures such as 1D-CNNs, RNNs/GRUs, dilated convolutions, or Temporal Transformers.
Comfortable with ensembling and cascade-style model architectures — combining multiple models’ outputs and evaluating latency/complexity vs. performance trade-offs.
Solid grasp of evaluation metrics for imbalanced, multi-class time-series classification/segmentation across hundreds of classes (precision/recall trade-offs, label noise, sensor placement variance).
Able to work independently with a large, configuration-driven codebase, respecting established architectural conventions and software design patterns.