Peeyush Dyavarashetty

About

AI/LLM practitioner who built a GPT-2-like language model from scratch at the University of Maryland using PyTorch and multi-GPU distributed training, with experiment tracking in Weights & Biases. As an AI Operations intern at ScaleUp360, delivered multiple production-style AI agent automations (Gmail classification and Fireflies-to-Claude workflows that extract and assign CEO tasks) and set up measurable evaluation using test cases and classification metrics.

Experience

Education

Key Strengths

• Built a GPT-2-like LLM from scratch in PyTorch, including architecture and loss function work
• Hands-on experience scaling training/validation across multiple GPUs (distributed training)
• Diagnosed and mitigated training slowdowns/stalls by separating training and validation workflows
• Implemented experiment tracking and visualization for model performance using Weights & Biases
• Designed and shipped multiple AI-agent automation workflows (Gmail classification; Fireflies-to-Claude task extraction for CEO)
• Established structured evaluation process for AI agents using test cases and metrics (accuracy/precision/recall) plus relevance of assigned tasks
• Effective iteration with a non-technical stakeholder (CEO) to translate business needs into AI agent capabilities

Reference Highlights

• Thorough in understanding requirements and details before proposing solutions
• Effective collaboration with non-technical stakeholders
• Clear, structured technical explanations (explains purpose of tools/services before solutions)
• Communicates rationale for choosing one tool over another
• Avoids repetitive explanations
• Contributed to taking an LLM-based product to production on AWS