Avinash N

Led a 15-member Tamil language team to drive quality evaluation for LLM responses in Tamil and English, improving team QC pass rates from ~34–50% to over 95% within four months while reducing average handling time. Conducted weekly QC reviews for 100+ tasks, assessing rating accuracy, search query formulation, source quality, justification strength, and evidence alignment. Evaluated citation relevance/necessity/alignment with model outputs, identified hallucinations and unsupported claims, onboarded new members across 10+ language teams, and collaborated with other leads to update evaluation guidelines.

Evaluated model-generated Python code across functional correctness, logical validity, constraint satisfaction, edge-case handling, test-case pass rate, readability, and exception handling. Classified code errors by type and severity including compilation/runtime failures and instruction-following gaps, and compared expected vs actual output mismatches. Evaluated AI-generated images against reference images for visual consistency, facial preservation, style accuracy, artifact detection, and prompt coherence, providing scoring rationales to guide model improvements. Annotated short-form videos by captioning on-screen text and labeling user actions with precise time-frame alignment to strengthen temporal understanding.

Built Python scripts to automate about 70% of data preprocessing tasks, reducing manual effort by approximately 15 hours weekly. Collaborated with other interns to develop a sentiment analysis model for customer feedback using NLTK projects. Developed a high-precision fraud transaction prediction model achieving 99% accuracy measured by AUC. Implemented a BERT-based model for tweet sentiment classification.