Whenever you're ready, walk me through how you'd approach this.
Discuss your experience optimizing model weights, quantization, and running inference on resource-constrained hardware.
Explain how you would reduce model size, memory use, and latency while preserving task quality. Include a practical implementation that compares a floating-point model with a quantized version, measures accuracy and runtime, and identifies when Core ML conversion, post-training quantization, or quantization-aware training is appropriate.