The last map-phase episode: we build the workshop and set the strategy. We construct the development environment the whole course uses, then make the case that shipped portfolio projects beat certificates in a tougher entry-level market.
Why Python won. Python is glue over compiled cores; NumPy brings C and Fortran power to Python by dispatching array math to BLAS/LAPACK. Proof of the hybrid design: SciPy is roughly 50% Python, 25% Fortran, 20% C (SciPy 1.0, Nature Methods; Array Programming with NumPy). Python 3.14 (Oct 2025) ships a supported free-threaded build, but pin 3.11 or 3.12 for ML since CUDA wheels lag (Astral).
Environments & packaging. venv+pip (baseline), conda/Miniforge (binary deps like CUDA), and uv (Rust, 10-100x faster, the 2026 default). Note the Anaconda licensing landmine: paid license required for 200+ employee orgs; prefer Miniforge/conda-forge. Lockfiles give reproducibility.
Notebooks. Great for exploration, terrible for git. Pimentel et al. analyzed 1.45M notebooks: only ~24% ran clean, only 4% reproduced results. Cure: Restart Kernel and Run All. Tools: nbstripout, jupytext, nbdime; marimo as a reactive alternative.
Git & GitHub. Commit code/configs/lockfiles, not data/weights/secrets (100MB cap). README in Problem-Solution-Impact format. Later: Git LFS and DVC.
Editors & hardware. VS Code dominates (73.6% in SO 2024); Cursor, Copilot, Claude Code are the AI trio. You don't need a GPU early; use free Colab (T4) and Kaggle.
Portfolio over credential. WEF Future of Jobs 2025: net +78M jobs by 2030; AI/ML specialists among the fastest-growing. ML Engineer pay ranges widely (Coursera). Entry-level is genuinely tough (IEEE Spectrum), which is exactly why a portfolio matters: 3-5 repos, one deployed demo, one write-up (Careery).
News: SpaceX/xAI's record $75B IPO (NPR); Moonshot's open-weight Kimi K2.7 Code; OpenAI retires GPT-5.2; Google's DiffusionGemma.