I am currently finishing a PhD focused on increasing the security of text classifiers against adversarial attacks. Adversarial attacks are when text inputs are subtly altered to mislead AI classifiers into making incorrect predictions, without the text input losing its original meaning or grammaticality. My research is on developing new training approaches to train generative NLP models to quickly generate adversarial examples for a victim text classifier. The training approach will depend on many factors, such the allowed knowledge of the attacker (from no knowledge to full access) to the design of the text classifier (what NLP task it serves, is it mono or multilingual, etc). After designing the techniques, I implement them with pytorch, comprehensively test their effectiveness, and write and present the results in research papers. My PhD has been financially supported by both the Australian Government and CSIRO's Data61. It is comprised of four papers, all currently under review, with the highlight probably being developing the first generative attack to target multilingual classifiers. I have been lucky enough to present my work both in showcases at UTS and at national conferences (ALTA 2022 & 2023). Some other side jobs include- Teaching assistant for undergraduate courses - Research assistant, completing project work to fill grants

Employee #1 at a machine learning/consulting hybrid startup. This role had two components:1. Prototyping, testing and implementing ideas for machine learning products and apps2. Sales, lead generation, marketing, blogging, community building and other general startup essentialsMost notable project: built prototype app where users could search for property using voice natural language queries, e.g. "Show me all one bedroom apartments within 500m of a train station".

Six month contract working across a range of projects, from recommendation engines to customer value segmentation.In this role I used a wide range of technologies, from the more traditional (R, Python, SQL) to the more production-related (Kubernetes, Docker, Google Cloud Platform) to the stranger (Unity). A recurring theme was making queries and scripts efficient on "big data", as Woolworths has a large number of transactions. Notable projects:- Built a recommendation engine to make personalised recommendations for BWS products (alongside three other data scientists)- Built machine learning pipelines and models for customer value segmentation- Built an virtual reality data visualisation that you could "walk around" and interact with

Data scientist on an business-internal consulting team. In this role I worked with stakeholders right across the business and tackled a wide range of projects. Our emphasis was on demonstrated business value: as a team, we aimed to create $15m deliverable value each year. This meant it was important to value practical solutions over theoretical ones and to present the work to people with a wide range of technical backgrounds. Notable projects: - Implemented a causal impact model to measure cannibalisation of beverages whena competing product was introduced. The model partially led to Caltex's incorporation of Boost Juice into many of its stores, netting large amounts of additional profit.- Redesigned and rebuilt an end-to-end geospatial machine learning pipeline foridentifying profitable locations to build Caltex petrol stations. The model incorporated twelve sources of data, including traffic, census and competitor locations, and it achieved a 50% increase inpredictive accuracy over its predecessor. - Designed and implemented a non-linear optimisation model for space allocation of retail goods. Further improved the model by allocating different space recommendations to different clusters of sites. The model was used to reallocate space across over the Caltex network

Data scientist for the Big Data and Enterprise Call Centre teams. This was a varied role. Day-to-day work was mostly writing SQL scripts on our Big Data cluster (which used Hadoop), but it also included creating intricate data visualisations with d3.js, and implementing R machine learning models.Notable projects - Built, alongside two other data scientists, a SQL model to identify network failure locations for the copper ADSL network. This was a predictive maintenance project and we used a variety of data sources, including high-volume network performance data. We then ran a pilot where we sent technicians to the predicted fault locations, who then verified if there was a fault or not. The pilot was successful and the program is on track to be rolled out nationally, potentially saving many millions of dollars. - Designed and built an interactive visualisation tool of the ADSL network's performance usingD3.js. Users could quickly identify network failures and the visualisation tool was adopted by the business. - Implemented a SVM (this was before transformers) to classify network fault reports as "valid" or "invalid". I created features from free text fields, tuned the model hyperparameters using cross-validation, and at the end, was able to retrospectively classify over 400,000 fault reports.

Data analyst working in network outage prediction. This role was mainly data exploration, analysis, visualisation and presentation to stakeholders. Notable projects: - Built topic classification model for Telstra's customer verbatims. I started by manually annotating over 3000 examples, before then training an ensemble model. I distilled and presented the results, and the work led to Telstra focusing more on problem resolution for customers, and less on other factors, such as giving them discounts- Built internal R package for simple heat-map construction, enabling other analysts to quickly analyse trending terms in customer outage reports- Statistical time series analysis of monthly network outages