Designed and built 3 multithreaded web services in flask and gunicorn that can store and rerun as many as 5000 historical models at the same time through remote containers managed by Kubernetes workers

Reduced model processing time by over 20% by empowering analysts to better utilize smart caching and parallel processing through serializing remote model-run contexts and aggregating all historical execution dependencies from cached outputs

Building image segmentation models to estimating the leaf coverage distribution in a garden and segment the leaves according to the plant types in the garden based on images taken from a top-down view camera.

Reduced training latency by 50% and improved model precision by 10% by building learnable PyTorch kernels in C++ that provide support for lower-bit aggressive quantization aware training on both CPU and GPU