Stakeholder interview (30 minutes)
Ask in this order; write answers into the spec for scripts/frame_check.py.
- What decision do you make today that this should improve? Who makes it, how often?
- How do you make it now? (This is the baseline.) How well does it work, and how do you know?
- If you had a perfect prediction, what would you do differently? How many cases could you handle per day/week?
- What exactly counts as the outcome? Where is it recorded? How long after the decision is it known?
- How far ahead do you need to know? How long does acting take?
- Which data do you look at when deciding? When is each piece of it updated?
- What happens when the prediction is wrong in each direction? Which mistake is worse, and by how much?
- Does acting change the outcome? (If yes: holdout, uplift.)
- Who else is affected? Any regulated or sensitive groups?
- What would make you stop using the model?
Spec fields read by frame_check.py
| Field | Type | Used by rules |
|---|---|---|
problem | string | context only |
decision, action | string | FR001 |
action_capacity | string | FR010 |
owner | string | FR013 |
unit_of_prediction | string | FR003 |
prediction_time | string | FR004 |
target | string | FR002, FR005, FR014 |
task_type | one of binary_classification, multiclass_classification, regression, forecasting, ranking, survival, uplift, anomaly, clustering | FR005, FR010, FR014 |
horizon_days, time_to_act_days, label_delay_days | number | FR005, FR006 |
label_is_proxy, proxy_risk | bool, string | FR011 |
features | list of {"name", "available_at"}; available_at in prediction_time, before_prediction_time, static, or a later moment | FR007 |
baseline | string | FR008 |
success_metric, guardrail_metrics | string, list | FR009 |
prediction_affects_label, holdout_plan | bool, string | FR012 |
Review checklist
Decision
- A named decision, owner and action
- Capacity per period
- Today's process is described and can be measured (baseline)
Prediction
- Unit of prediction = entity at a moment
- Prediction time and latency budget
- Horizon longer than the time to act
- Training population built with the scoring eligibility logic
Label
- Written as a query
- Outcome window fully observed for every training row
- Label delay known; monitoring plan accounts for it
- Proxy gap documented
- Noise estimated on a double-labelled sample if humans label
Features
- Every source has an "available at" time
- No outcome-side, final-state or all-time aggregate columns
- As-of logic for slowly changing data
Evaluation
- Primary metric at the operating point
- Guardrails
- Validation split mirrors deployment (time-based / grouped)
- Sliced metrics for key segments and protected groups
Risks
- Feedback loop: holdout and action logging
- Selection bias in labels
- Regulatory and explanation requirements
- Retraining cadence and drift monitoring
Red-flag phrases
| Phrase | Ask |
|---|---|
| "Use AI to..." | Which decision changes? |
| "Predict churn" | Within what window, for whom, scored when? |
| "We'll figure out how to use it later" | Then stop until you do |
| "We have all the data" | When is each field written? |
| "Accuracy of 95%" | What is the base rate? What does the current rule get? |
| "Real-time" | What latency, and is the data real-time too? |
| "The model will learn what matters" | From which label, measured how? |
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.