AI Data Labeler / RLHF Evaluator (Travelite project)
Designed retrieval pipelines for a travel assistant, focusing on intent labeling and evaluation of conversation quality. Labeled ambiguous model outputs and failure modes, feeding back improvements for model tuning. Implemented RLHF-style preference ranking and instruction-feedback cycles. • Evaluated chat responses for natural language understanding. • Labeled user queries by intent and correctness. • Iteratively refined model outputs through annotation. • Produced feedback and quality signals to enhance AI retrieval.