Multimodal Sentiment Analysis for E-commerce Optimization
I led a high-precision data labeling initiative to train a multimodal LLM for a major retail client. The project involved categorizing 50,000+ user-generated reviews by sentiment (Positive, Negative, Neutral) and intent. I performed granular text annotation alongside image classification to ensure the model correctly identified sarcasm and context-dependent visual cues in product feedback. We maintained a 98% inter-annotator agreement (IAA) through rigorous QA loops and custom labeling ontologies.