Data & ML quickstart

1. Look at the data first (15 min): shape, dtypes, missing values, duplicates, target distribution, 5 raw rows. pandas: df.info(); df.describe(); df.isna().mean(). Write down what each column means and what leaks the answer.
2. Define the question & metric before modelling (regression → MAE; classification → F1/PR-AUC when imbalanced; ranking → top-k hit rate). Tie it to the user's decision.
3. Baseline first: majority class / mean / last value / simple rule. Then logistic or linear regression, then gradient boosting (lightgbm/xgboost/sklearn). Stop when it beats baseline meaningfully.
4. Split correctly: by time for time series, by group/user for repeated entities; no leakage (fit scalers/encoders on train only). Keep a final holdout untouched until the end.
5. Pre-trained over trained: for text/images/audio use hosted or open models (embeddings, classification, OCR, speech) rather than training from scratch.
6. Explain: feature importances / SHAP / example errors. Show 3 failure cases honestly.
7. Ship it: save the model (joblib) or call the API; expose via the backend (backend-api); cache predictions; pin library versions in requirements.txt.
8. Visualize for the pitch: one clear chart per message (use a plain bar/line, labelled axes, highlighted takeaway).
9. Scale story: batch vs streaming, retraining schedule, monitoring drift, feature store — talk, don't build.

Keep notebooks out of main unless needed; put reproducible logic in src/ scripts. Large data → don't commit; document the download step in docs/STACK.md.

Done when: a script reproduces the metric from raw data in one command and the number is on a slide with its baseline next to it.