Competitions
Flutter, Dart, TensorFlow Lite, ONNX, PyTorch
- Built a 100% offline Flutter-based mobile AI system to classify rice quality on low-end Android devices by optimizing a ConvNeXt Vision Transformer via ONNX graph surgery and FP16 quantization, reducing model memory footprint by 50% and tile input size by 75%.
- Prevented OOM crashes on MediaTek Helio CPUs using Dart background isolates and ARM big.LITTLE core tuning, enabling real-time predictions from 12MP images without UI lag on 6GB RAM devices.
Python, PyTorch, Google Earth Engine
- Built a hybrid satellite time-series classification system to predict crop types from Sentinel-2 imagery by optimizing seasonal pattern learning using PatchTST transformer embeddings fused with vegetation indices (NDVI, NDRE, SAVI).
- Achieved robust classification of irregular multivariate field data using a soft-voting ensemble of Random Forest, SVM, and Logistic Regression, achieving 0.7473 Macro F1.
