Euan Turner

Exploring the potential of GPU-accelerated interval analysis to accelerate the bounding of optimisation problems.

Ongoing

Supervised by Dr. Calvin Tsay and Phil Sosnin

Block diffusion language models process many tokens in parallel, inflating the number of activated experts. This challenges existing expert offloading methods for auto-regressive methods that rely on sparse expert activation.

My dissertation co-designs a token-aware, saliency-guided expert utility metric with diffusion-specific expert caching and execution. This preserves 91% of the model’s baseline performance, while yielding a 7.86-10.6x improvement in TPS over existing works.

Supervised by Dr. Hongxiang Fan

Worked on improving inertial odometry models; specifically their ability to generalise or adapt to new users or activity contexts.

Developed a strategic roadmap for integrating self-supervised learning into the model pipeline, grounded in an extensive literature review.

Conducted a detailed analysis comparing proprietary and open-source datasets, identifying key factors contributing to model performance issues. For example, a subset of externally collected data were most likely measured from within pockets, a severely more challenging scenario.

Significantly improved the efficiency of hyper-parameter turning by incorporating KerasTuner.

Investigated failure modes of inertial odometry models on specific users, identifying a core mismatch between velocity training losses and metrics.

Decreased QEMU live upgrade times by a factor of 10, and eliminated network traffic demands. Persisted virtual machine memory and virtual device state in place by adding file descriptor passing mechanisms between the VMs.

Saved developers minutes on of setup to debug QEMU, by automating host configuration of required tools during deployment.

Collaborated with open source community on upstream changes, developing a new perspective on version control as a logical presentation for review.

Reached early prototype of solution within two weeks, rapidly acclimating to several complex legacy codebases.

Won the US and Europe intern hackathon out of 20 teams, having take responsibility for presentation of the project to global team leaders.

Developed a scalable pronunciation training app by integrating ensemble ML models (Allosaurus and Whisper) for real-time phoneme inference.

Engineered a Dockerized FastAPI backend with an NGINX load balancer to auto-scale inference microservices across multiple languages.

Optimized model performance by adopting an instance-per-call strategy for Allosaurus, reducing latency and improving throughput.

Leveraged AWS CloudFront CDN to deliver phoneme recordings, cutting global response times by over 67%.

Made extensive use of Scrum ceremonies, with sprint planning and retrospectives proving particularly helpful in refining workflows and ensuring project alignment.

Awarded the David Howarth Group Project Prize for best overall group project

Prototyped a web application that would save UK farmers on the order of thousands of pounds a year with environmental scheme compliance. Centralised task and field management together with evidence compilation, removing manual efforts to collate and distribute the required information.

Effectively utilised the agile methodology, working closely with farmers nationwide from project definition through to prototype. Validated incremental design choices at each stage, incrementally creating intuitive workflows, leading to high stakeholder satisfaction by the end of the project.

Developed a Next.js web app in TypeScript, utilising the Prisma ORM for database interaction.

Recognised as the 4th best project out of 56 in the year.

Leading the operations committee for a non-profit supporting university athletics clubs in London and the South-East.

Led a committee of 16 students, co-ordinating the delivery and financial management of two years of events.

Restored financial viability of the club, from a deficit of £2000 to a surplus of £4000 over two years, by taking direct oversight over all event finances.

Returned Hyde Park Relays, a 76 year-old event, to profitability, ensuring its feature for the foreseeable future. Led the improvement of organisational procedures to satisfy the Royal Parks event licensing team.

Represented Imperial at regional and national competitions.

Recognised with Imperial Athletes’ Colours for “outstanding contribution to Imperial Athletes and the wider Imperial community”.

Delivered small group tutorials on Discrete Maths, Logic and Reasoning about Programs.

Assistant students during programming practicals in first and second year.