Atul Deshpande

LLM Efficiency Benchmarking Framework

Engineered parameter-efficient fine-tuning pipeline for GPT-2, Phi-2, LLaMA-2 and Mistral-7B using cloud GPU infrastructure.

• Built automated evaluation pipeline using BLEU, ROUGE-L, BERTScore
• Curated 2K technical QA dataset
• Demonstrated superior sample efficiency of modern architectures

Industrial RAG System for Maintenance Analysis

Designed a Retrieval-Augmented Generation system for analyzing industrial maintenance logs.

• Metadata-driven retrieval for contextual answers
• Machine ID and issue-based filtering

Technical Question Answering with LLMs

• Built RAG pipeline using Mistral-7B-Instruct
• MiniLM embeddings with FAISS vector search
• Evaluated RAG vs Fine-tuned systems