Feasibility scorecard template
| Dimension | Status | Evidence | Mitigation / owner |
|---|---|---|---|
| Framing | ok | decision, unit, label window agreed (link) | - |
| Volume | risk | 82 positives, EPV 7.5 | restrict to 8 features; pool 6 more months (data eng, 2 wks) |
| Labels | ok | label rate stable by month; kappa 0.78 | - |
| Signal | ok | best OOF AUC 0.69 (tenure) | - |
| Leak suspects | blocking until resolved | refund_issued AUC 0.95; cancellation_reason missingness AUC 1.0 | drop; confirm write times |
| Availability | risk | support tickets only in warehouse (24 h lag) | accept 24 h staleness |
| Measurability | risk | AUC CI half-width 0.06 | report recall@400 with CI; extend test window |
| Value | ok | lift ~$10k/week at prec@400 0.40 vs 0.25 | - |
Verdict rule: any blocking -> NO-GO (or "GO after X"); any risk -> GO WITH RISKS with named mitigations; all ok -> GO.
Pilot stages
Stage 0: desk check (1-3 days)
- Framing check, feasibility scan, sample-size arithmetic, value model.
- Output: scorecard and the decision to spend on stage 1.
Stage 1: offline prototype (1-2 weeks)
- Time-based holdout mirroring deployment (train on months 1-12, test on 13-15).
- Baseline (current rule, dummy, single feature) and one simple model (logistic regression or a small GBM).
- Report the primary metric at the operating point with a CI.
- Kill if: model lift over the baseline is within the CI, or the lift is below break-even.
Stage 2: shadow mode (2-4 weeks)
- Score live traffic without acting. Log features, scores, latency.
- Check: feature availability rate, latency p99, score distribution vs offline, drift.
- Kill if: key features missing for > X% of rows, latency over budget, or live scores look very different from offline ones.
Stage 3: controlled test (4-8 weeks)
- Randomise: act on the model's top-K in the treatment arm, the baseline's top-K in control (or a random holdout with no action, to measure the action's effect).
- Primary metric: the business outcome (saves, fraud $ prevented), with guardrails.
- Decide with a pre-registered analysis plan and sample size.
Kill criteria examples
- "Recall@400 under the model is not at least 20% higher than the rule, with the CI excluding 0."
- "Fewer than 90% of scored rows have the top-5 features populated in shadow mode."
- "p99 latency > 50 ms after optimisation."
- "Uplift in saves per 400 calls < 5 in the controlled test."
- "Fairness: false positive rate ratio between regions outside 0.8-1.25."
Go/No-Go write-up template
# Feasibility: <project>
Verdict: GO WITH RISKS - signal and value are adequate; two probable leaks must be removed and volume is thin.
## Framing
<one-sentence framing>; label = <definition>; horizon <h>; capacity <K>.
## Scorecard
<table above>
## Data needed
<rows/positives/months>; collection time <x weeks> (sample_size.py rate ...).
## Value
Baseline <metric>, model target <metric>; lift <$/period>; ceiling <$/period>; break-even prec@K <x>.
## Pilot plan
Stage 1 by <date>, stage 2 by <date>, kill criteria: <list>.
## What would change the answer
<for NO-GO: new data source, longer window, rules first>Common NO-GO outcomes and the alternative
| Reason | Alternative |
|---|---|
| Too few positives | rules + label collection; anomaly detection; widen the event definition |
| No signal | new data sources; qualitative research into what drives the outcome |
| Signal only after the event | reframe to a later decision point, or a different decision |
| Capacity too small for value | improve the action (better offer) before the targeting |
| Labels unreliable | fix the labelling process; measure kappa again |
| No consumer system | integration work first; the model can wait |
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.