Deveshree Mohile

Assessed conversational outputs for Uber’s support interactions serving millions of riders and drivers by designing structured rubrics and labeling frameworks across 15+ intent categories. Collaborated with applied science and product teams to convert evaluation findings into actionable insights, identifying failure patterns and edge cases to improve response quality and reduce policy violations while increasing factual accuracy. Analyzed large-scale GenAI evaluation datasets using SQL to quantify performance trends and support data-driven decisions for model iteration and deployment readiness. Implemented API development practices to streamline data ingestion and retrieval processes for evaluation datasets and contributed documentation for reproducibility and knowledge sharing within the GenAI Evaluations team.

Assessed conversational outputs for Uber support interactions at very large scale by designing structured rubrics and labeling frameworks to evaluate response quality, factual accuracy, and policy adherence across 15+ intent categories. Collaborated with applied science and product teams to convert evaluation findings into actionable insights, identify failure patterns and edge cases, and refine prompts—reducing policy-violation rates by 87% and improving factual accuracy scores by 75%. Analyzed large-scale evaluation datasets by translating them into SQL-based trends to inform model iteration and deployment readiness. Implemented API development practices to streamline data ingestion and retrieval for evaluation datasets, improving integration efficiency with internal tools. Contributed documentation of evaluation methodologies and data processing workflows to support reproducibility and knowledge sharing; supported problem-solving by diagnosing data inconsistencies and collaboratin

Guided 40+ graduate students in mastering Bayesian estimation, maximum likelihood estimation, and minimum mean square error methods by providing targeted feedback and connecting statistical estimation theory to practical machine learning through bias-variance tradeoffs and real-world model evaluation.