At Sri Lanka Telecom Headquarters, I collaborated on an FTTH fault prediction project, preprocessing and analyzing 28M+ telecom records to support model training and evaluation. I applied PatchTST for multi-step fault prediction and benchmarked and tuned model performance.
In my plant disease recognition research project, I developed lightweight models and used knowledge distillation and quantization-aware training to reduce model size by up to 50% while maintaining accuracy. I also deployed the optimized Potato disease model in a Flutter app with ONNX Runtime for offline on-device inference.
