VARUN NEGI

• Built and fine-tuned deep learning models for classification and generation tasks, achieving 92% classification accuracy and reducing inference latency by 25% through model quantization and batching. • Designed end-to-end ML pipelines using Python, TensorFlow, and AWS, cutting model deployment time from days to under 2 hours via automated CI/CD workflows with Docker and GitHub Actions. • Developed RAG (Retrieval-Augmented Generation) pipelines using LangChain and ChromaDB, enabling semantic document search with sub-200ms query latency across 50K+ document corpora. • Fine-tuned LLMs (GPT-3.5, LLaMA 2) using LoRA on domain-specific datasets, improving task-specific accuracy by 18% over base models. • Applied Prompt Engineering techniques (chain-of-thought, few-shot, structured output) to optimize LLM responses, reducing hallucination rate by 35% in production. • Contributed to LLMOps practices: model versioning, eval tracking with MLflow, and automated regression testing on 200+ prompt-response pairs per release.

• Built and maintained full-stack web applications using JavaScript, Python (FastAPI/Flask), MySQL, and MongoDB, serving 500+ daily active users. • Optimized backend query performance and API response times, reducing average page load by 25% and cutting database query cost by 40% through indexing and caching with Redis. • Integrated 10+ third-party REST APIs and delivered 3 major feature releases on schedule within agile sprint cycles. • Conducted code reviews and mentored 2 junior developers, improving team PR merge velocity by 20%.