Paul Tune Email and Phone Number

Leading the Marketing Optimization team in the Marketing Technology (MarTech) side of things. In particular, our team concentrates on building tools and product features that would help a marketer to speed up their work in creatives and value optimization. Some examples of our work are* an Apple Search Ads keyword and bid optimization system based on survival analysis and contextual bandits,* predictions of user value which feeds in a Mix Media Model to save us millions of dollars per year, * ad copy title generation tools using LLMs, and* an automated layout resizing model to scale ad asset creation for enterprise-scale marketing campaigns.

Canva seeks to empower the world to design, and serves a base of over 20 million monthly active users in more than 100 countries. My responsibilities include:improving search and relevance for our users via automated search elevations and machine-learned boost functions, building recommender systems to provide personalised recommendations for users,building models to target users at Canva better to upgrade to our premium product,improve Canva's retention rate, building propensity models and architecting systems that optimise our marketing channels, andproductionising assorted models in the service of simplifying the design process for the user.Other tasks include improving the user data extraction and transformation pipeline, and generating reports for determining the effectiveness of products and marketing campaigns.

I worked on implementing a deep learning-based general object classification and detection algorithm to run at scale on distributed systems. In detail, my responsibilities include:Optimise the image processing pipeline to utilise GPUs fully, and lower latency in each step along the pipeline,Training and experimenting with new deep learning models for fast person detection,Preparing datasets and debugging deep learning models when they underperform,Solving object recognition problems such as detecting occluded objects and finding them in bad lighting conditions (example, in the night and under bad weather conditions), andProvide input on the API design.Image Intelligence is still in its infancy, so like my fellow colleagues, I wear several hats, including taking on DevOps duties occasionally.

Developed a motion detection algorithm and hashing scheme to run on distributed systems at Cammy. Introduced metrics for accessing object recognition algorithms at Cammy.

Develop statistical inference methods for recovering and synthesizing traffic matrices, which are important for the operations maintenance of networks.

Worked on the development of new algorithms for data reduction of flows traversing high speed, national carrier routers, as part of a grant awarded by the Institute for a Broadband Enabled Society (IBES), part of the National Broadband Network initiative in Australia. Compared existing algorithms to assess their efficiency in obtaining information from these flows. The end goal of the project is to develop algorithms that enable efficient real time diagnostics of these high speed networks.

Worked on the statistical performance of sampling and randomized data structures for the collection of traffic from the Internet. The traffic we look at are from national carriers, the routers that transport traffic from one end of the country to the other, or to overseas. Showed that the current methodology i.e. packet sampling, employed in Cisco's NetFlow products, is inadequate to perform accurate estimation of flow statistics. The work done is theoretical (involving estimation theory) with some practical work done on actual traffic traces, to validate our conclusions.Also working on recovering sparse signals, a special class of signals that is ubiquitous in nature. For example, take the standard information captured by a digital camera. Once compressed in JPEG, more than 90% of information (even as high as 97%!) is discarded. This begs the question: if we are going to collect very little information, why not do it at the start, rather than discarding excess data at the post-processing stage? This is where compressed sensing comes in. Apart from deriving performance bounds, being the minimum number of measurements required when the signal is contaminated by noise, we also apply this new paradigm to communication systems.

Worked on a project under A/Prof Peter Farrell at the University of Melbourne in association with the Ludwig Institute for Cancer Research (LICR). The project involves the detection of cancer cells via bioluminescence, as certain markers used can be stimulated to emit light once in contact with luminol. Based on the intensity of the light, we estimate the number of cancer cells in the sample. The challenge was that the light emitted was very low in intensity and thus, we designed a circuit to detect the light. We then perform curve fitting to estimate cancer cell counts in a sample.

Prototyped security system using a microcontroller with GSM capabilities. Involved circuit design and programming the device with a variant of C.