Luke Martin

• Extended the functionality of state of the art diffusion based protein models to work within our existing generation system. Building on an opensource model, adding in conditional generation options whilst maintaining the prior performance. Built in Pytorch. • Led the design of a novel protein optimization platform, bringing together numerous protein property prediction, interaction, and numerical models. This technology is a core feature of Ordaos’ offerings and has been able to improve binding by 10x, with resilience to target evolution.

• Implemented portions of a generative multi-modal protein attention model using PyTorch, predicting a number of protein properties, 3D structure, and sequence. • Curated the data for, implemented, designed, and trained an antibody interaction model to predict the binding strength (log Kd) value of an antibody and target pair. • Fine-tuned the log Kd model on in-house data, which was able to rank from a set of prospective antibodies, placing 57 successful binders within the top 70 candidates, 46 of which had high affinity.

• Built a paper ingestion system that pulled all new papers from arXiv into a database as an Azure microservice. • Worked with a colleague on a convolution vision model to detect cancer in prostate tissue slides. • Using BERT embeddings on paper abstracts, built a paper recommendation service that recommended the most relevant papers from our database of papers.

• Designed, implemented, and validated models of the SARB (South African Reserve Bank) cash-flow system using Tensorflow. • Improved the percentage accuracy over the prior existing model that predicted the movement of cash in regions across the country by up to 40%. • Assisted colleagues in building a model to optimize the routing of SARB cash in transit vehicles.