With an MSc in Language & Speech Technology from UCL, Tatiana combines linguistic expertise with data science and machine learning. As a Data Scientist at Wyser, she builds evaluation pipelines that measure how well large language models actually perform: structured, criteria-based judges, error analysis across models and languages, and DSPy to optimise prompts systematically rather than by hand. 

Her work sits at the intersection of NLP and applied machine learning. She benchmarks open-weight and commercial models, builds the schemas that shape how AI extracts and structures information and tracks model behaviour with tools like MLflow to keep evaluation honest and reproducible. Working closely with engineers, researchers and customers, she turns questions like “is this model good enough?” into measurable, defensible answers. 

Her work also feeds into Wyser’s wider research into responsible AI, making sure the tools built on language models are transparent, inclusive, and genuinely usable by the organisations and communities they serve.

What drives me?

“Most models are built and tested on a narrow slice of how people actually communicate. I want to close that gap.” 

In three words

Curious, analytical and patient

What’s one misconception about AI you’d love to change? 

People assume that because a model was trained on huge amounts of human-generated text, it understands the world the way we do. It’s learned patterns in how humans communicate, but that’s not the same as lived context. A model doesn’t know what it’s like to be tired, to code-switch mid-sentence, or to hedge because you’re unsure how something will land. Finding where that gap shows up in real outputs is what my evaluation work is for. 

What excites you most about the future?

Speech and language technology that works for everyone, including the accents and languages it currently gets wrong. 

Outside of Wyser… 

Outside of work, I’m usually reading, lifting weights, or over-analysing a chess position.