Ship better on-device AI/ML faster

Developed an internal cross-platform ML inference library;• Supports constructing arbitrary ML pipelines from JSON configurations.• Leverages modern neural network hardware on edge-devices.• Supports intermingling CPU, GPU, and NPU workloads.• Generates optimal asynchronicity for complex execution graphs.

Deep Render are designing a deep learning video compression pipeline.My role was mostly split between development of Deep Render's custom implementation of the rANS (Range Asymmetric Numeral Systems) entropy coder written in C++ and converting our research ML pipeline into a viable product on consumer devices.• Performance analysis and optimisation of our entropy coder, pertaining to runtime, energy usage and bitstream size.• Keeping our entropy coder up-to-date with our overall pipeline requirements.• Facilitating feature additions that require information to pass from our encoding pipeline to our decoding pipeline through injection of metadata into the bitstream.• Interfacing with researchers on model design decisions, particularly regarding the potential runtime performance impact of proposed additions/changes.• Porting of research pipeline for consumer device architectures (eg. Snapdragon/Apple chipsets), making use of NPUs and GPUs

• Performance analysis and optimisation for external researchers as part of EuroHPC, generally scientific simulations in C++ for distributed architechtures• Development of a C++ webserver (Deimos) for heirarchical storage management backend• Conference presentation on development of Deimos.• Design of API for heirarchical storage management solving data management for exascale architectures• Development and maintainance of GoLang RestAPI handling worklow management for large-scale execution (Lexis)• Integration of workflow-related features into the Lexis front-end; including visualisation of arbitrary workflow topologies• Preparation of material and delivery of a number of HPC-related courses