Certificate

# ===== Skills checklist =====
# After the ML track you should be able to:
# [x] Load, inspect, clean, and split a tabular dataset
# [x] Ship a dumb baseline before any model
# [x] Build a sklearn Pipeline with preprocessing + model
# [x] Cross-validate with the right split (stratified / group / time)
# [x] Detect + prevent data leakage
# [x] Pick the right metric for the problem (ROC-AUC, PR-AUC, MAE, NDCG)
# [x] Compare XGBoost / LightGBM / CatBoost for tabular SOTA
# [x] Save + load model artifacts via joblib / mlflow
# [x] Track experiments (mlflow, W&B)
# [x] Deploy a model behind a small FastAPI / Flask service
# [x] Monitor for drift (KS test, PSI) in production
# [x] Plan retraining cadence + retrain pipeline

# ===== Bookmark =====
# - https://scikit-learn.org/stable/user_guide.html
# - https://xgboost.readthedocs.io
# - https://www.lightgbm.io
# - https://huggingface.co/learn (for NLP)
# - https://goodfellow.io (Deep Learning book)
# - https://huyenchip.com/dmls-book (Designing ML Systems)
# - https://wandb.ai (experiment tracking)

# ===== Portfolio project (8-16 hours) =====
# Build an end-to-end ML feature with a public-facing demo:
# 1) Pick a real problem (churn, fraud, price prediction, recommendation)
# 2) Source a dataset (Kaggle, UCI, public APIs)
# 3) EDA notebook: distribution plots, missing values, target balance
# 4) Baseline: always-predict-majority OR mean; report its metric honestly
# 5) Pipeline: ColumnTransformer + StandardScaler + OHE + your model
# 6) Train with cross-validation; pick the model that beats baseline
# 7) Evaluate on held-out test set ONCE
# 8) Save artifact via joblib
# 9) FastAPI endpoint that loads the artifact and serves predictions
# 10) Streamlit / Gradio demo UI hosted on Hugging Face Spaces
# 11) README that explains framing, baselines, metric, results, limitations

# Bonus:
# - mlflow tracking with hyperparam sweep
# - shadow-deploy against the baseline for a week
# - drift monitoring with Evidently
# - A/B framework + a writeup of an experiment

# ===== What 'good' looks like =====
# - Test metric is reported with a confidence interval
# - Baseline is named, evaluated, and beaten by a stated margin
# - The pipeline is reproducible: one command trains the model from raw data
# - The README explains WHY this model was chosen, not just WHAT it does
# - The hosted demo handles missing values without crashing
# - Documented limitations (e.g. 'works only on customers with > 6 months of data')

# ===== Common mistakes to avoid =====
# - Showing a model with no baseline (looks impressive, says nothing)
# - Cherry-picked metric (e.g. accuracy on imbalanced data)
# - Leaky feature pipeline (mean computed on full data, not per-fold)
# - Random KFold on time-shaped data
# - Skipping the inference test on edge cases
# - 'It works on Kaggle' but never tested in a real serving environment

# ===== Next steps =====
# - Pick a specialisation: NLP, computer vision, time series, RecSys, MLOps
# - Read the canonical papers in your specialisation
# - Contribute to an open-source ML library
# - Build features end-to-end at work: 'productionised ML' is the rare skill
# - Mentor someone through this track; teaching cements the concepts

# ===== Self-test =====
# If you can:
# 1) Reframe a vague business problem as a labelled ML task with the right metric
# 2) Ship a baseline + model + serving endpoint in a weekend
# 3) Explain leakage to a teammate and design a CV strategy that avoids it
# you have completed the track. Ship the portfolio piece and call it done.

# ===== Track wrap-up =====
# 90% of production ML problems are tabular and beaten by gradient boosting on
# well-prepared features. The four habits that matter:
# - Baseline first
# - Right split for the data shape
# - Pipeline that captures preprocessing so test/train cannot diverge
# - Evaluate honestly with a meaningful metric
# Get those right and the rest is technique.

# /certificate/ml