An Edge AI framework for real-time arrhythmia diagnosis from ECG signals. Combines 1D-CNN feature extraction with LoRA parameter-efficient adaptation (< 100 KB), weight fusion, and INT8 dynamic quantization for sub-millisecond on-device inference.
Continuous cardiac monitoring on wearable and edge devices is bottlenecked by strict power, memory (< 50 MB), and latency (< 10 ms) constraints. Furthermore, high inter-patient physiological variability causes generic deep learning models to misclassify benign morphologic variations as arrhythmias.
Architected a decoupled MLOps lifecycle: a base 1D-CNN trained with inter-patient splits is personalized with LoRA on the first 150 patient beats (< 100 KB adapter). Weights are merged via `merge_and_unload()` and quantized into INT8 ONNX models with zero PyTorch runtime dependency.
Engineered the full pipeline from DVC data versioning and MLflow tracking to causal bandpass filtering, 1D-CNN design, LoRA fine-tuning, INT8 quantization, and containerized FastAPI inference with automated health checks.