Petro Milan

Petro Milan is a senior solutions architect at NVIDIA specializing in large-language-model inference and agentic AI. His career spans rocket-engine simulations, scientific instruments and enterprise AI, with a consistent focus on making computationally demanding models practical.

From propulsion research to enterprise AI

In 2019, Milan participated in Argonne National Laboratory’s extreme-scale computing program, working on parallel multiphysics simulations for liquid rocket engines. He earned a doctorate in aerospace engineering and a master’s degree in computational science and engineering from Georgia Tech in 2021. His doctoral dissertation applied deep learning to propulsion-related flow simulations.

At SambaNova Systems, Milan became first author of a 2022 study of adaptive machine learning for X-ray free-electron lasers, comparing graphics processors with reconfigurable dataflow hardware for retraining models as experimental conditions change. By 2024, he was a principal AI engineer at SambaNova, where he led hands-on development of Llama 3 applications at the AI Engineer World’s Fair. He subsequently joined NVIDIA as a senior solutions architect.

• Inference speed shapes product design. Milan evaluates responsiveness through throughput, time to first token and total response time; sufficiently fast generation enables interfaces that revise answers while users edit prompts.
• Model-specific prompting prevents avoidable failures. Correct control tokens and conversation templates help instruction-tuned models produce bounded, relevant answers instead of unnecessary output.
• Enterprise retrieval requires end-to-end engineering. His workflows combine document parsing, configurable chunking, E5 embeddings, ChromaDB retrieval and grounded generation, allowing organizations to query changing internal information without continually retraining models.
• Hardware should match the workload. Embeddings can run locally or on accelerated infrastructure, while document processing and storage remain separate from inference; complex PDF tables may require richer extraction tools than straightforward documents.