How I used Microsoft Olive's SelectiveMixedPrecision and ModelBuilder passes to quantize Qwen2.5-1.5B-Instruct to INT4 — with strategic INT8 overrides — and deployed it on iOS using ONNX Runtime GenAI and Flutter.
How I exported Whisper Base from Hugging Face, quantized it to INT8 with onnxruntime.quantization, and built the full inference pipeline in Dart — mel spectrogram, tokenizer, autoregressive decoder, token streaming — without sherpa-onnx or any abstraction layer.
How I deployed Whisper Tiny INT8 natively on iOS using sherpa-onnx and Flutter Isolates — no cloud, no latency, no data leaving the device — as part of building a local-first personal health record app.
A deep dive into ONNX — the open format and runtime that decouples model training from deployment, enabling the same model to run on CPUs, GPUs, mobile chips, and browsers without rewriting a line of inference code.
How ONNX-based optimization reduced medical image segmentation inference time by 95%, plus practical edge deployment patterns for mobile and healthcare AI systems.
From million-year tectonic cycles to seconds of earthquake rupture — why geodynamic simulation is computationally brutal, why building training data is even harder, and what machine learning can realistically do about it.
How Physics-Informed Neural Networks (PINNs) embed PDEs directly into the learning process — and why they are transforming earthquake science, seismic imaging, and fault mechanics.
How Artificial Neural Networks and Random Forests can predict earthquake rupture propagation with >81% accuracy.