Step by step
- Inventory and measure the rules. For a labelled period, compute each rule's hit rate, precision and incremental recall, and the whole rule set's precision/recall at its current volume. This is the bar.
- Check label provenance. If outcomes were observed only for cases rules flagged (reviews, blocks), the labels are biased toward what rules already catch. Plan a random review sample (e.g. 1-2% of unflagged cases) to observe outcomes elsewhere.
- Split policy from prediction. Separate rules that encode policy (must stay) from rules that are heuristics for risk (candidates for replacement).
- Build features the rules use, and more. Rule conditions are good features; keep them so the model can at least match the rules.
- Offline comparison. Same holdout, same capacity: rule precision@K vs model precision@K. Include
learned rules (
rule_learner.py) as an intermediate option. - Shadow mode. Score live traffic with the model; log decisions both ways; review a sample of disagreements with domain experts. Expect the model to find some cases rules miss and vice versa.
- Champion/challenger. Route a small, randomised share of traffic to model + policy guardrails. Compare business outcomes with a pre-registered analysis.
- Ramp up and retire. Increase traffic in steps; retire heuristic rules whose incremental recall drops to ~0 under the model. Keep a rules-only fallback path for outages.
- Keep learning signals. Keep the random review sample and analyst overrides; they feed retraining and detect drift.
Selection bias from rule-generated labels
Rules flag 5% of traffic -> analysts review flagged cases -> labels exist mostly for flagged cases
Model trained on these labels learns "what analysts confirmed among rule hits"Consequences: the model can look excellent offline and still miss fraud types the rules never flagged.
Mitigations:
- Random exploration sample with outcomes observed (review or delayed chargeback).
- Train on outcomes that do not depend on review (e.g. chargebacks for approved transactions).
- Weight explored samples by inverse selection probability.
- Report metrics separately on the exploration sample.
Rollback plan
| Trigger | Action |
|---|---|
| Precision on the challenger falls below rules for 2 consecutive days | route traffic back to rules |
| Feature pipeline outage | automatic fallback to rules |
| Score distribution shift (PSI > 0.25) | alert; pause ramp-up |
| Complaint or false-decline spike | revert and investigate |
Pitfalls
| Pitfall | Effect | Fix |
|---|---|---|
| Dropping policy rules into the model | policy violations when the model disagrees | keep policy rules as hard guardrails |
| Training only on rule-flagged outcomes | model inherits blind spots | exploration sample |
| Big-bang switch | no comparison, no rollback | shadow, then small traffic share |
| Retiring rules without measuring incremental recall | losing coverage of rare patterns | measure under the model before deleting |
| No logging of which layer decided | audits impossible | log rule ID / model version per decision |
| Analysts lose the ability to react to new attacks | slow response | keep an emergency-rule path |
anomaly-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.