Sofiia Lobanova

AI governance research

• Researching structured transparency for safety evaluations and how AI systems become embedded in social decision-making processes.

Testing AI Incentives, with Simon Goldstein, Peter Salib, and Yonathan Arbel

• Co-first author of AI Revealed Preferences. Led the Quora workstream end to end, building a dataset of 25K+ LLM responses across 20 models with 15 engineered features.
• Proposed and implemented a Bradley-Terry / Elo-style ranking method for forced-choice experiments on real-world Quora data.
• Found a consistent revealed aversion to “uncomfortable-truth” tasks across all 20 tested models, a behavioral signal relevant to deception and scheming evaluation.

Aligning AIs Via External Incentives, with Peter Salib, Simon Goldstein, and Cole Niblett

• Extending the revealed-preferences and incentive-sensitivity research toward tighter experimental isolation and a publishable output.

Production NLP on a 27B-parameter Russian-language LLM

• Own parallel NLP tracks from dataset design through production deployment and reliability monitoring on a system processing about 300K social-media posts a day at peak.
• Built evaluation and annotation infrastructure that raised label quality above 90% on previously unreliable categories; introduced MLflow and Weights & Biases.

Clinical-reporting product, parallel project

• Architected and shipped an LLM pipeline for OCR, voice transcription, multi-step summarisation, and document generation.
• Built disagreement detection and a writer/editor prompt architecture for reliability in a medically sensitive setting.
• Cut report turnaround from roughly two working days to about an hour, and parsing errors from about 30% to under 1%.

NLP, computer vision, and geospatial ML for international clients

• Worked across applied NLP, satellite imagery, and geospatial ML for research covering more than 1,000 cities and projects including a Fortune Global 500 company and UNDP.
• Delivered social-media analytics, crosswalk detection at F1 ≈ 0.93, building segmentation, and a low-resource-language research chatbot.

City-scale mobility and transport analytics

• Built passenger-flow, demand, parking-policy, EV-siting, and carsharing analyses on the city mobility data lake.

Mixed-methods research across finance, e-commerce, and healthcare

• Conducted 200+ in-depth interviews across 20+ projects, combining experimental design with quantitative behavioral analysis.