Rahul Kumar Sharma

Write, review, debug, and optimize Python scripts generated by large language models, identifying logical, syntactic, and performance issues and producing production-quality reference code. Evaluate LLM outputs through RLHF-style workflows by ranking responses, correcting model outputs, and providing structured preference signals. Design prompts, evaluation rubrics, and expert reference solutions for supervised fine-tuning and LLM evaluation pipelines. Assess model reasoning, factual correctness, and instruction-following while collaborating with cross-functional AI teams on annotations and code evaluations.

Designed and implemented a Python NLP pipeline using NLTK, VADER, Hugging Face Transformers, and PyTorch to classify financial news headlines by sentiment, achieving more than 92% accuracy. Performed text preprocessing and feature engineering, including tokenization, stopword removal, and lemmatization. Optimized GPU-accelerated inference with CUDA and cuDNN, reducing latency by 40% and supporting near-real-time processing of more than 10,000 headlines. Used FinBERT, Matplotlib, and Seaborn to analyze sentiment trends and volatility signals.

Created prompts and evaluated LLM-generated responses in RLHF workflows, comparing outputs, assigning quality ratings, and providing detailed feedback. Reviewed, debugged, corrected, and optimized AI-generated Python code across diverse programming tasks, identifying logical errors, syntax issues, inefficiencies, and readability problems. Produced reference solutions, documented model failure cases and reasoning gaps, and delivered annotations, preference rankings, and technical feedback to support large-scale AI training and evaluation projects.