At Cohere Labs Open Science Community, I co-lead the analysis subteam for a multicultural riddles benchmarking project studying how LLMs handle culturally grounded reasoning tasks.
The analysis focuses on factual accuracy, hallucination behavior, topic-level patterns, and cross-cultural reasoning failure modes. I help coordinate analysis workflows for a 15-person subteam inside a 100+ researcher open-science initiative, aligning model benchmarking, human evaluation, and synthesis outputs.