• Developed Generalised Linear Mixed Models to predict click-through rate (CTR) and conversion rate (CVR) • Fitted models using variational inference on billions of data points to improve training efficiency and remove bias from gradient updates; built the training algorithm such that it could distributed to exploit data parallelism and model sparsity

• Trained Generative Adversarial Networks (GANs) to predict 3D human pose estimates from one RGB image• Used image data augmentation techniques to improve model robustness to ``in the wild'' images • Incorporated depth estimation networks to produce global 3D pose predictions, used by Kinect for Azure

Course staff for CS109 - Probability for Computer Scientists - Winter 2017 (Jan 2017 -- Mar 2017).Course staff for CS110 - Principles of Computer Systems - Spring 2017 (Apr 2017 -- Jun 2017).Course staff for CS110 - Principles of Computer Systems - Fall 2017 (Sep 2017 -- Dec 2017).Course staff for CS234 - Reinforcement Learning - Winter 2018 (Jan 2018 -- Mar 2018).Course staff for CS110 - Principles of Computer Systems - Spring 2018 (Apr 2018 -- Jun 2018).Duties included:- Holding office hours- Running labs- Grading- Code reviewing- Writing exam questions- Writing lecture notes

• Improved delivery of alarms by removing the necessity of sound files to be stored on devices.• Added several features that improve the scalability of the alarm systems, for example, enabling audio files stored on DynamoDB to be used, rather than files having to be stored locally on the device.

Committee member of Churchill Spring Ball 2016 as head of security. An event with over 900 guests, over £66k budget and £60k revenueResponsibilities included: • Booking and liaising with a security company• Organising fencing and stage barriers• Arranging a wristband system• Hiring and organising the radio communications system• Managing the security team on the night• General setup and clean up for the event

A small start-up (at the time I worked there), working on a restful “geo-spatial” database system written in Java, on top of Google’s App Engine. The geo-spatial database provided a RESTful API for clients to interact with. A corresponding website provided usage statistics, billing information and more. Technologies used included Javascript, (Facebook’s) react and GoLang.Examples of work:• Co-ordinated a distributed cluster of servers to test the maximal throughput of the system.• Mathematically modeled the system to identify bottlenecks between constituent modules.• Developed an eventually consistent local cache for server instances, increasing maximal throughput by approximately 5000req/sec (approx 2-3x what could be handled before) under a realistic ratio of query types.