Phiona

At Cantine / Phiona I led the research and development of deep learning models that infer psychological and behavioural traits from audio. It is a hard problem: labelled data is scarce and labels are subjective. To get around that, I used self-supervised pre-training and representation learning on large unlabelled audio datasets, then fine-tuned Transformer models on smaller annotated sets.

I also built multimodal and generative models combining audio with behavioural signals, using GANs and other generative approaches for audio synthesis and data augmentation, and designed the full ML pipeline from data collection and annotation through training, statistical evaluation and benchmarking to production deployment.

The work was carried out in collaboration with researchers at University College London (UCL) across audio ML, generative AI and computational psychology. I built and led a team of three engineers, introducing CI/CD, automated testing and experiment tracking to take research models into production.