Petro Junior Milan is a senior solutions architect at NVIDIA specializing in large-language-model inference and agentic AI. His career spans aerospace engineering, scientific machine learning, specialized computing hardware, and enterprise generative AI.
Milan completed a doctorate in aerospace engineering and a master’s degree in computational science and engineering at Georgia Tech in 2021, following a 2020 internship at Argonne National Laboratory. His doctoral research applied deep learning to computationally intensive propulsion and fluid-dynamics problems, including neural-network approaches to calculating real-fluid properties.
He joined SambaNova Systems in 2021 to work on AI for science and coauthored research on adaptive machine learning for X-ray free-electron lasers. Selected for the National Academy of Engineering’s U.S. Frontiers of Engineering symposium in 2023, he advanced to principal AI engineer at SambaNova, where his responsibilities expanded into prompt engineering, retrieval-augmented generation, fine-tuning, and agentic workflows. He subsequently joined NVIDIA as a senior solutions architect.
Technical priorities
- Scientific machine learning for physical systems. Milan has applied neural networks to fluid-property calculations, propulsion simulations, and scientific instruments requiring fast model adaptation.
- Enterprise retrieval grounded in real documents. At AI Engineer World’s Fair 2024, he led the practical portion of a SambaNova workshop, building a document question-answering pipeline with PDF ingestion, configurable chunking, E5-large-v2 embeddings, ChromaDB, semantic retrieval, and Llama 3. Related examples are available in SambaNova’s company-owned AI Starter Kit.
- Inference performance beyond token throughput. He distinguishes generation speed from time to first token and total response latency, connecting those measurements to responsive chat, live drafting, and interactive prompt development.
- Deployment details that shape application quality. His implementation priorities include model-specific prompt formatting, generation parameters, document-parser selection, retrieval configuration, optional reranking, and whether embeddings run on conventional processors or specialized hardware.
