Anastasia Tsukanova

I am the leader of a team specializing in quality control for AI projects (computer vision, LLMs), responsible for establishing the quality assurance strategy, communicating the standards to developers, and supporting them with AI assessment and XAI tools such as TracIn, SHAP, t-SNE. We examine various data, model and application aspects of AI quality, such as data consistency, coverage, balance, model accuracy and robustness, application privacy, security and fairness. In LLM projects (GPT, LLaMa, Claude), I contributed as the main prompt engineer, responsible for the identification of the rare cases where the output did not correspond to our expectations (e.g., hallucination, endless repetition, timeout error), and adjusting the prompts accordingly or developing workarounds. In computer vision projects, I led my team to identify the model’s critical areas for improvement and deeply understand the origins of that behavior.

Created custom client-oriented linguistic resources meeting established quality criteria for a range of natural language processing applications. Designed hard tasks to train and test LLMs compared to native speakers in terms of language accuracy, coherency, tone, and cultural sensitivity and awareness. Developed robust linguistic resources and models for the Russian natural language understanding unit of in-car voice assistants, investigating the origin of bugs, and identifying opportunities to improve the application.

Developed a deep neural network-based articulatory speech synthesizer, did extensive work to automatically extract articulatory information from real-time MRI data.

Modeled articulatory information extracted from static MRI data.

Mathematical modeling of distributed search in a peer-to-peer network.