Arzaan Ul Mairaj

Data Scientist @Turing

CA, US

Created task scenarios to evaluate Gemini’s agentic behaviour in workspace automation, improving the model’s ability to understand user intent and perform reliably in practical workflows.• Built custom datasets and applied RLHF to improve Google Gemini’s Chain‑of‑Thought reasoning because Google required better handling of complex data analysis and multi‑step tasks for its Data Agents, greatly enhancing its reasoning capabilities in structured scenarios and improvedoutput quality.• Conducted in‑depth benchmarking of Gemini against other LLMs, identifying performance gaps in reasoning, verbosity, and coherence to understand limitations and optimize model performance in comparison to leading alternatives, directly informing fine‑tuning strategies and improved RLHF training design.• Designed curated datasets focused on real‑world analytical tasks because existing datasets lacked the complexity needed for robust model evaluation, leading to improved predictive accuracy and clearer outputs from Gemini in practical applications.