Where the effort goes (typical tabular ML project to first launch)
| Work | Share | Notes |
|---|---|---|
| Data access, label definition, point-in-time features | 35-50% | the most underestimated part |
| Baselines, modelling, validation, error analysis | 15-25% | |
| Serving integration, feature pipelines in production | 15-25% | batch is much cheaper than real-time |
| Shadow mode, online test, analysis | 10-15% | calendar time dominated by the test duration |
| Monitoring, documentation, handover | 5-10% | model card, runbook, dashboards |
Estimation heuristics
- Data first: if any key source has no owner or no history, add 2-4 weeks.
- Real-time vs batch: real-time features and serving add 1.5-3x engineering effort.
- New label pipelines (surveys, human labelling) add the labelling calendar time plus quality checks.
- Online test duration comes from the sample-size calculation, not from the project plan: e.g. 6-8 weeks for weekly decisions.
- Regulated domains: add legal/model-risk review cycles (often 2-6 weeks each).
- Add 30% contingency; ML projects rarely run short.
Milestone design
Each milestone states an artefact and a checkable exit criterion:
| Milestone | Artefact | Exit criterion examples |
|---|---|---|
| Data + labels | training table, label SQL, data dictionary | label reviewed by the data owner; leakage scan clean; row counts reconcile with finance |
| Baseline | baseline table | current process, dummy, simple model on a deployment-like split, with CIs |
| Candidate | model, validation report | beats baseline at the operating point; error analysis by segment; calibration if needed |
| Shadow | live scoring logs | feature availability >= 98%; latency within budget; live vs offline score PSI < 0.2 |
| Online test | test analysis | pre-registered primary metric significant at the planned power |
| Launch | dashboards, alerts, runbook, model card | on-call owner named; rollback tested |
Staffing patterns
| Project size | Team |
|---|---|
| Small (batch, existing data, one decision) | 1 DS + 0.3 DE + product owner 0.1 |
| Medium (new features, batch serving, A/B test) | 1-2 DS + 1 DE + 0.5 ML engineer + PO |
| Large (real-time, many teams, regulated) | 2-3 DS + 2 DE + 1-2 MLE + analyst + PO + model-risk reviewer |
Phase gates
Write the gate decision options explicitly: continue, pivot (change framing/scope), or stop.
| Gate | Continue if | Pivot if | Stop if |
|---|---|---|---|
| After baseline | simple model beats rule clearly | signal exists but metric/decision mismatch | nothing beats dummy |
| After candidate | target met with CI | target met only for some segments | gap to target > half the planned lift |
| After shadow | live matches offline | features unavailable for a segment | systemic train/serve skew |
| After test | primary metric significant | positive but below target | no effect or harm |
Timeline example (weekly batch model)
Week 1-2 data access, label SQL, point-in-time features
Week 3 baseline table <- gate 1
Week 4-7 modelling, validation, error analysis <- gate 2
Week 8-11 batch pipeline, shadow mode <- gate 3
Week 12-17 randomised test <- gate 4
Week 18 launch, monitoring, model cardanomaly-detection-model-builder
Build and evaluate anomaly detection for tabular, sensor and transaction data with few or no labels: define the anomaly type (point, contextual, collective), choose detectors (robust z-score/MAD and IQR rules, Mahalanobis with robust covariance, k-NN distance, LOF, Isolation Forest, one-class SVM, PCA reconstruction / autoencoders, forecasting residuals for time series), combine them by rank, explain each alert by its deviating features, set thresholds from an alert budget or the score distribution, and evaluate with precision@k, PR-AUC and recall on known incidents, or detector agreement when no labels exist. Includes pure-Python detectors (robust z, Mahalanobis, kNN, Isolation Forest, ensemble) and an evaluation script.
catboost-model-builder
Build CatBoost models that use its strengths correctly: raw categorical columns as cat_features (ordered target statistics, no manual encoding), text_features, Pool construction with NaN-safe categories, ordered boosting and has_time for temporal data, symmetric (oblivious) trees, overfitting detector with eval_set and use_best_model, losses for classification, regression (Quantile, Poisson, Tweedie) and ranking (YetiRank), class weights, monotone constraints, feature importance (PredictionValuesChange, LossFunctionChange) and SHAP values, GPU training and model export. Plans cat/text/numeric columns from a CSV and generates a runnable training script, and checks parameter sets for the conflicts CatBoost rejects (iterations vs n_estimators, Bayesian bootstrap with subsample, multiple class-weight options, depth > 16, use_best_model without eval_set).
categorical-encoding-selector
Choose and implement the right encoding for every categorical column given the model family and the data: one-hot (with infrequent-level grouping and handle_unknown), ordinal with an explicit order, cross-fitted target encoding (scikit-learn TargetEncoder, CatBoost ordered statistics, category_encoders inside CV), frequency/count encoding, feature hashing, native categorical support in LightGBM, HistGradientBoosting, XGBoost and CatBoost, prefix levels for codes (zip3, NAICS-2), and embeddings. Measures cardinality, rare-level share, unseen-level rate on a time-ordered holdout and target signal per column, emits the ColumnTransformer, and quantifies target-encoding leakage (in-sample vs out-of-fold vs holdout AUC).
classification-model-selector
Choose the right classification algorithm for tabular, text or mixed data and prove the choice with a fair bake-off: logistic regression (L2, L1, elastic net), linear SVM, naive Bayes, k-NN, RBF SVM, decision trees, random forest, extra trees, HistGradientBoosting, LightGBM, XGBoost, CatBoost, explainable boosting machines (EBM/GAM), points scorecards and MLPs, matched to data size, width, categorical and high-cardinality share, missingness, class imbalance, text, and constraints such as interpretability, monotonic constraints, calibrated probabilities, latency budgets and tiny deployable models. Ranks candidates with starting hyperparameters and cautions, then generates a bake-off script running every candidate on identical CV folds with per-fold metrics, fit time and prediction latency.
clustering-algorithm-selector
Choose the right clustering algorithm and check that clustering makes sense at all: cluster tendency (Hopkins statistic with a correlated-features caveat, nearest-neighbour isolation), then k-means / MiniBatchKMeans, Gaussian mixtures with BIC, agglomerative (Ward, average, complete, single) with dendrograms, DBSCAN, HDBSCAN, OPTICS, spectral clustering, BIRCH, mean shift, k-prototypes and k-modes for mixed or categorical data, and Gower distance, matched to data size, expected shapes, noise, varying density, known k, soft memberships, hierarchy and streaming. Covers preprocessing (scaling, encoding, dimensionality, distance metrics), validation (internal indices, bootstrap stability, external labels, business profiling) and assigning new points.
collaborative-filtering-implementer
Implement collaborative filtering recommenders that beat popularity: item-based kNN with cosine, Jaccard or asymmetric similarity, shrinkage, neighbourhood size, positive-only or mean-centred rating signals and 'because you liked X' explanations; implicit ALS with confidence weighting, BPR pairwise ranking and the closed-form EASE model in plain NumPy, tuned on a temporal validation split by recall@k and NDCG@k and refit on all data; plus user-kNN, explicit matrix factorisation, the implicit/LightFM libraries, hyperparameter ranges, cold-start fallbacks, scaling with sparse matrices and approximate nearest neighbours. Produces top-k lists in the format of the recommendation evaluator.