Research lead content perception and narration, 02/23-present- Image narrator an asynchronous API which has used a variety of Large Language Models (Flan, GPT4, GPT4 vision (image, video)), and Visual Question Answering Models (BLIP2, Instruct BLIP, GPT4v image) to better understand visual content, narrative, and humour. - Prototyped the content narrator for video content using pre-release GPT4 vision text and video tools, showcased in Microsoft Ignite, crafting the initial narrative, script, presentation, and solution prototype and AI question-answering conversation. Demonstrates what could be achieved when combining video intelligence with the performance brain. - AI-Driven Audio Description Tool which provides in-flight audio descriptions which maintain best practice accessibility descriptions whilst aligning with brand and marketing strategy. Tool covers segment detection, video description and analysis, caption creation evaluation and iteration. - Interested in the interplay between visual perception, guardrails, and storytellingTech lead for performance brain // content intelligence 04/22-02/23- Led the technical discovery, proof of concept, and development to MVP for the ground-breaking "Performance Brain," predicting, explaining, and optimising marketing performance.- Managed a cross-functional team of data scientists and engineers, ensuring collaborative efforts between technical, product, and commercial teams.

An Intel funded computer science studentship position to identify and develop solutions that identify and reduce machine learning failures in autonomous vehicle perception and planning systems. My research focused on using asymmetric neural network ensemble architectures to improve machine vision performance. This position was created to build on the innovations I made in my MSc thesis that applied the safety-critical computing concepts of defence in depth and design diversity to machine learning. My research focuses on applying concepts of asymmetry and design diversity to ensemble design, data, parameter search, and training. information to convolutional neural network ensembles within a realistic autonomous vehicle software stack embedded within a photorealistic urban driving simulation. Project implemented using ROS, Autoware, Carla, OpenCV, Docker, Python, and TensorFlow.Withdrew from PhD studies in September 2022 due to professional opportunities and family commitments.

Appsbroker is one of the largest Google Cloud-only Agile Systems Integrators and Managed Service Providers in EMEA.I work in cross-functional agile teams where I am responsible for building machine-learning-based models that deliver value for clients using modern data science development tools such as Python, Docker, Git and Unix. Provided pricing optimisation and sales modelling for a global FMCG client, where I used linear models, EPOS information, and time-series data to predict weekly sales across different countries, markets, and product lines to integrate with promotion and pricing optimisation algorithms developed by the wider team. This lead to increases of over £3m potential profit from smarter pricing in just one of the sixty-plus product categories across key European markets. I designed and ran various experiments to test the importance of product placement information demonstrating that product placement information from crowdsourced information providers was insufficient for their purposes. Projects built using Python, sci-kit learn, statsmodel, and GCP.

I was responsible for the design, development, and documentation of proof of concept machine learning solutions for clients. My biggest contribution was the design, development and delivery of a data mining and explainability tool for action-state time-series data for DSTL to identify, analyse, and explain strategies used by agents in competitive multiagent reinforcement learning simulations for applications of cyber warfare technologies in military aviation scenarios. The tools used include ensemble learning, clustering, SHAP, information theory, and natural language processing. This significantly reduced the FTE and time required to analyse, understand and communicate in natural language how a cyber arms race scenario could unfold. It also provided insights into the actions and strategies that adversaries could use as mitigations against their opponent. Other projects include a counter ai system for naval warfare scenarios as part of DIEM’s red mirror product that was funded by DSTL’s Intelligent Warship research program, an OCR solution for an international postage company to extract parcel metadata, and internal visualisation, reporting and analysis products for DUCHESS, an information elicitation chatbot, to inform its product development roadmap and scrum backlog. Tools included JupyterLab, Azure, Python, PostgreSQL, NTLK, SpaCY, GCP compute engine, GCP storage, Seaborn, and Vader.

This was an internship as part of my MSc course where I worked in collaboration with Adelard, a safety-critical computing consultancy, and TIGARS, an international research group for autonomous automotive vehicles.The purpose of this internship was to research how to make autonomous vehicles safer using the concepts of defence in depth and design diversity.My first innovation was to use ensemble learning for convolutional neural networks, the basis for machine learning perception systems, as the mechanism to introduce defence in depth. My second innovation was to introduce design diversity through the introduction of asymmetries to the specification, optimisation, and training of individual ensemble member models. The impact of this research was that it increased the increased safety of an autonomous vehicle in a toy scenario across several measures. Firstly, increased resilience to changes in the operational design domain, meaning that when the car operated on a new track, at new times of days, it retained higher performance than other methods analysed. Secondly, it outperformed the evaluation metrics of comparable symmetric neural network ensembles as well as highly optimised individual convolutional neural networks, and VGG with transfer learning, meaning that it was an empirically better solution. The third is increased robustness to changing operational input data, meaning that it overfitted less to the training data and as such could generalise better to new unseen observations and had less sudden reactions to changes in the road. A key resulting benefit from the asymmetric neural ensemble is that they require fewer processing chips in a vehicle compared to standard k-fold neural network ensembles due to pipelining of classification then regression predictions, reducing potential costs of production. Implemented using DonkeyCar, Linux, Python and TensorFlow on a Keras backend.

Automated processes covering the collection, validation, cleansing, and reporting of finance MI data using visual basic, excel, access, SQL, and Oracle Essbase. This reduced the average FTE required to complete MI month-end reporting by 70%, and time to delivery decreased by 60%. Projects built using visual basic, excel, access, SQL, and Oracle Essbase.

Member of portfolio and risk analytics team that provided comprehensive data services, portfolio analytics, credit risk model development and stress testing for the UK bank.My biggest achievement was the design and development of a predictive RAROC pricing system that was used by all relationship managers and lending divisions within corporate and retail banking to assess and price all new on and off balance sheet products and services. It leveraged all of the risk metrics developed by the model development team. It improved the banks capital allocation process by enabling it to more accurately and consistently apply risk-adjusted return on capital evaluation all product lines, increasing the speed that relationship managers can price new business, enabled relationship managers to have more data-driven and interactive conversations with clients and business stakeholders, and helped the bank achieve its object of becoming a model-driven organisation. I built the system using Excel, VBA, PowerView, R, SharePoint, and SharePoint workflows with some assistance from IT and mentoring from the head of risk analytics. Built and analysed a portfolio of multiple linear regression models, Monte Carlo simulations, and distribution fitting techniques to calculate crucial portfolio financial and risk indicators to stress test the bank portfolio as part of the ICAAP process. This enabled the bank to understand if and how it might survive a variety of adverse economic scenarios set by the Bank of England. This required expert statistical modelling and economic knowledge, the interrogation of the centralised data warehouse, as well as R and SAS model development. Developed a variety of supervised and unsupervised machine learning models in R and SAS to monitor and predict changes in capital losses and changes in portfolio risk metrics.Contribution recognised through Risk Leadership Award from UK Managing Director.

A contracted Product Analyst where I provided comprehensive analysis of ŠKODA market position and product performance to enable product managers to better price and equip vehicles so that they can balance profitability KPIis against market share, and to help communication managers and digital managers prioritise, customise, and select their marketing campaigns. Biggest achievement was leading a range rationalisation project identifying poor performing derivatives to enable optimisation of pricing, product specification, and stock management,

Strategic marketing that included market research, interrogation of CRM database and financial messaging enabled Toybox to better strategically plan its engagement policy with umbrella groups and its supporter base.Analytical investigation into UK church charitable giving which enabled the charity to launch a targeted fundraising and marketing campaign, increasing supporter count, supporter awareness of issues that the charity is trying to address, and revenue.Designed, developed and maintained a database for institutional grant givers along with a bulk grant application tool. This increased team visibility of sources of funding decreased the effort required to apply to low chance funding sources and built a knowledge repository to improve productivity and content for high-value applications. This increased revenue and productivity of the grant applications team.