I’m building a virtual DeepseekV3 environment that emulates Jet Nano hardware for research and development on machine-learning models. The goal is to give my team a sandbox where we can move seamlessly from data preprocessing and feature extraction through model training, evaluation, deployment, and monitoring—without touching the physical board until we are ready. Here’s what I need: • A reproducible simulation that mirrors Jet Nano’s CUDA-enabled GPU, memory constraints, and I/O. • Containerised tool-chain (PyTorch, TensorRT, cuDNN, etc.) with scripts that cover the full life-cycle: preprocessing, training, hyper-parameter sweeps, evaluation metrics, and a mock-deployment stage that tracks resource usage and latency. • Clear documentation so any teammate can spin up the environment, run the sample pipelines, and swap in new datasets or model architectures. Acceptance criteria • End-to-end demo shows a small dataset flowing through preprocessing → trained model → virtual deployment with real-time monitoring. • All code runs on a fresh Ubuntu VM with one command. • Performance metrics inside the sim closely match published Jet Nano benchmarks (within 10 %).