• Design distillation framework for multimodal perception models.

• Reduce training cost for larger multipurpose and foundational models.

• NAS (neural architecture search) as framework, define search spaces for model applications.

• Various: Triton custom kernels. Kernels for sparse formats. ARM based devices. Custom silicon.

• Working closely with teams doing HW architecture, ML compiler, ML framework and ML researchers, publishing / open sourcing own research.

• Develop edge TPU optimized model architectures. Open source to Google's TensorFlow Model Garden.

• Building new backbones for visual perception using CNN and attention-based models. Building heads for semantic and instance segmentation and detection.

• Post-training model inference optimization

• Fitting large models to memory by optimizing sharding patterns, spills, and fills.

• General-purpose TPU computing, sparsity fit for TPU, quantization methods to optimize performance.

• Recommendations for ML hardware, compiler, and runtime design.

• Analytical performance and power modeling, automated memory layout,

• Standard model transformations in the compiler,

• Post-silicon performance validation, and software stack optimization.

As a technical leader, I was responsible for hiring and managing a new team for our North Carolina facility. We focused on four main areas: algorithmic work in computational geometry using mostly C++, systems development for CAM software using C++, PY, on-premises, and for the cloud, exploratory analysis and business decision support using various types of data available, and future product design and physical modeling using FEA.

Applications of machine learning and image processing.

Created team of 10+ people

Software that does automatic treatment target generation for internal treatment planning service

Software project of research nature (getting general availability of solution requires long systematic investigation). Successfully brought it from idea to production. Team of 4-6 development engineers