Sanjana Shellikeri, PhD

Senior Consultant, BLDS LLC

I'm a Senior Consultant at BLDS LLC, where I work on the statistics of fairness in machine learning. A lot of what I do lately involves large language models, where the question of what fairness even means is more contested than it sometimes looks from the outside. My clients are typically in credit, insurance, employment, or healthcare, where a model's output has tangible consequences for someone's life and where the wrong answer eventually shows up in front of a regulator.

I came to this work through a long detour. I trained as a neuroscientist and speech-language pathologist. I did my MSc and PhD in Speech-Language Pathology at the University of Toronto, then spent four years as a postdoctoral fellow at the University of Pennsylvania's Frontotemporal Degeneration Center, building digital speech biomarkers for ALS, frontotemporal dementia, Alzheimer's, and primary progressive aphasia. Around 25 peer-reviewed papers came out of that decade, including work in Neurology and Alzheimer's & Dementia, and one paper that the National Institute on Aging covered in a press release in 2024.

The move from clinical research into AI fairness made sense to me because the daily work is closer than it sounds. Whether the language being analyzed is a clinical speech recording or a stream of LLM outputs, the job is similar: extract an interpretable signal from messy text or audio, separate that signal from artifacts of how the data was collected or generated, and make claims about it that hold up to a skeptical reader. My current projects include building retrieval-augmented generation systems, analyzing embedding-based employment models that match candidate resumes to job descriptions, and training a small language model to detect discriminatory speech turn by turn so that each phrase can be flagged or denied at the moment of generation. The data look different from what I was working with in clinic. The methodological habits transfer almost directly.