#!/usr/bin/env python3
"""Render an ML project scoping document (Markdown) from a JSON spec.
Every section of a good ML scope is generated from the spec; anything missing is written as a
visible `TODO(owner)` so gaps cannot hide. Print a complete example spec with --template.
Sections: summary, decision and users, success criteria (primary metric, target, baseline,
guardrails, online metric), data and labels, approach (phased, simplest first), in scope / out of
scope, milestones with exit criteria, risks with mitigations, team and costs, go/no-go criteria,
open questions.
Usage:
python scope_doc.py --template > scope.json
python scope_doc.py scope.json > SCOPE.md
python scope_doc.py scope.json -o SCOPE.md --strict # exit 1 if any TODO remains
Exit codes: 0 rendered (no TODOs, or not --strict), 1 TODOs remain with --strict, 2 usage error.
"""
import argparse
import json
import sys
TEMPLATE = {
"title": "Churn retention targeting",
"owner": "Data Science - Retention",
"date": "2026-11-01",
"status": "draft",
"summary": "Rank active subscribers weekly by risk of cancelling in the next 30 days so the retention team "
"calls the 400 most at-risk customers.",
"problem": "Voluntary churn is 2.1% per month; the retention team calls customers whose contract ends this "
"month, which misses most churners.",
"decision": "Which 400 customers the retention team calls each Monday",
"users": ["Retention team leads (consume the list)", "Retention analysts (review weekly)"],
"success": {
"primary_metric": "recall@400 of customers cancelling within 30 days",
"target": ">= 0.45 on the latest 3-month time holdout",
"baseline": "0.30 for the current contract-end rule on the same holdout",
"guardrails": ["complaint rate of called customers <= 1.5%", "recall@400 per region >= 0.35"],
"online_metric": "saves per 400 calls vs a 5% untreated random holdout over 6 weeks",
},
"data": {
"sources": [
{"name": "billing.subscriptions", "owner": "finance", "freshness": "hourly", "available_at_prediction": True},
{"name": "support.tickets", "owner": "support", "freshness": "daily", "available_at_prediction": True},
],
"label": "voluntary cancellation within 30 days after the Monday scoring date",
"history": "24 months of weekly snapshots, ~50k active customers per week",
"known_issues": ["cancellation_reason is written after cancellation (excluded)"],
},
"approach": [
{"phase": "Baseline", "work": "current rule, dummy, logistic regression on a time holdout"},
{"phase": "Model", "work": "LightGBM with point-in-time features; calibrate; pick K=400"},
{"phase": "Shadow", "work": "score live for 4 weeks without acting; check drift and latency"},
{"phase": "Test", "work": "randomised test: model top-400 vs rule top-400, 6 weeks"},
],
"in_scope": ["weekly batch scoring", "voluntary churn", "consumer plans"],
"out_of_scope": ["involuntary churn (failed payments)", "real-time scoring", "offer personalisation"],
"milestones": [
{"name": "Baseline table", "date": "2026-11-15", "exit": "rule, dummy and logistic regression measured"},
{"name": "Model ready", "date": "2026-12-13", "exit": "recall@400 >= 0.45 offline with CI"},
{"name": "Shadow complete", "date": "2027-01-17", "exit": "feature availability >= 98%, PSI < 0.2"},
{"name": "Test readout", "date": "2027-03-07", "exit": "go/no-go decision"},
],
"risks": [
{"risk": "calling customers changes their labels (feedback loop)", "likelihood": "high",
"impact": "medium", "mitigation": "5% random untreated holdout; log calls"},
{"risk": "support tickets arrive with a 24 h delay", "likelihood": "medium", "impact": "low",
"mitigation": "use ticket counts up to the previous day only"},
],
"team": [{"role": "Data scientist", "fte": 1.0}, {"role": "Data engineer", "fte": 0.5},
{"role": "Retention lead (product owner)", "fte": 0.1}],
"costs": "about 4 person-months to the test readout; serving on the existing batch platform",
"go_no_go": ["offline recall@400 >= 0.45 with the 95% CI above the rule",
"online: +5 saves per week vs the rule with the CI excluding 0"],
"open_questions": ["Can the retention team handle 500 calls in December?"],
}
class Doc:
def __init__(self):
self.lines, self.todos = [], []
def add(self, *lines):
self.lines.extend(lines)
def val(self, v, what):
if v in (None, "", [], {}):
self.todos.append(what)
return f"`TODO(owner): {what}`"
return v
def render(spec):
d = Doc()
v = d.val
d.add(f"# {v(spec.get('title'), 'title')}", "",
f"| Owner | Date | Status |", "|-------|------|--------|",
f"| {v(spec.get('owner'), 'owner')} | {v(spec.get('date'), 'date')} | {spec.get('status', 'draft')} |", "",
"## Summary", "", v(spec.get("summary"), "one-paragraph summary: decision, prediction, value"), "",
"## Problem and decision", "", v(spec.get("problem"), "the business problem with a number"), "",
f"**Decision this model changes:** {v(spec.get('decision'), 'decision')}", "",
"**Users:**", "")
users = spec.get("users") or []
d.add(*([f"- {u}" for u in users] or [f"- {v(None, 'who consumes the predictions')}"]), "")
s = spec.get("success") or {}
d.add("## Success criteria", "", "| Item | Value |", "|------|-------|",
f"| Primary metric | {v(s.get('primary_metric'), 'primary metric at the operating point')} |",
f"| Target | {v(s.get('target'), 'target value and evaluation split')} |",
f"| Baseline | {v(s.get('baseline'), 'current process measured on the same split')} |",
f"| Guardrails | {'; '.join(s.get('guardrails') or []) or v(None, 'guardrail metrics')} |",
f"| Online metric | {v(s.get('online_metric'), 'how success is confirmed after launch')} |", "")
data = spec.get("data") or {}
d.add("## Data and labels", "", f"**Label:** {v(data.get('label'), 'label definition with window')}", "",
f"**History:** {v(data.get('history'), 'rows, time span, entities')}", "",
"| Source | Owner | Freshness | Available at prediction time |", "|--------|-------|-----------|------------------|")
srcs = data.get("sources") or []
if not srcs:
d.add(f"| {v(None, 'data sources')} | | | |")
for src in srcs:
avail = src.get("available_at_prediction")
d.add(f"| {src.get('name', '?')} | {src.get('owner', '?')} | {src.get('freshness', '?')} | "
f"{'yes' if avail else 'NO' if avail is False else v(None, 'availability of ' + src.get('name', '?'))} |")
if data.get("known_issues"):
d.add("", "**Known data issues:**", "", *[f"- {i}" for i in data["known_issues"]])
d.add("", "## Approach", "", "Simplest credible option first; each phase has an exit criterion in Milestones.", "")
phases = spec.get("approach") or []
d.add(*([f"{i}. **{p.get('phase', '?')}** - {p.get('work', '')}" for i, p in enumerate(phases, 1)]
or [f"1. {v(None, 'phased approach (baseline first)')}"]), "")
d.add("## Scope", "", "| In scope | Out of scope |", "|----------|--------------|")
ins, outs = spec.get("in_scope") or [], spec.get("out_of_scope") or []
if not outs:
outs = [v(None, "out-of-scope items (say what you will NOT do)")]
if not ins:
ins = [v(None, "in-scope items")]
for i in range(max(len(ins), len(outs))):
d.add(f"| {ins[i] if i < len(ins) else ''} | {outs[i] if i < len(outs) else ''} |")
d.add("", "## Milestones", "", "| Milestone | Date | Exit criterion |", "|-----------|------|----------------|")
ms = spec.get("milestones") or []
if not ms:
d.add(f"| {v(None, 'milestones with dates and exit criteria')} | | |")
for m in ms:
d.add(f"| {m.get('name', '?')} | {v(m.get('date'), 'date for ' + m.get('name', '?'))} | "
f"{v(m.get('exit'), 'exit criterion for ' + m.get('name', '?'))} |")
d.add("", "## Risks", "", "| Risk | Likelihood | Impact | Mitigation |", "|------|------------|--------|------------|")
risks = spec.get("risks") or []
if not risks:
d.add(f"| {v(None, 'risks (leakage, labels, adoption, feedback loops)')} | | | |")
for r in risks:
d.add(f"| {r.get('risk', '?')} | {r.get('likelihood', '?')} | {r.get('impact', '?')} | "
f"{v(r.get('mitigation'), 'mitigation for: ' + r.get('risk', '?'))} |")
d.add("", "## Team and costs", "")
team = spec.get("team") or []
d.add(*([f"- {t.get('role', '?')}: {t.get('fte', '?')} FTE" for t in team] or [f"- {v(None, 'team and FTE')}"]))
d.add("", f"**Cost estimate:** {v(spec.get('costs'), 'effort and run cost')}", "", "## Go / no-go criteria", "")
gng = spec.get("go_no_go") or []
d.add(*([f"- {g}" for g in gng] or [f"- {v(None, 'kill criteria written before the work starts')}"]), "")
oq = spec.get("open_questions") or []
if oq:
d.add("## Open questions", "", *[f"- {q}" for q in oq], "")
return "\n".join(d.lines).rstrip() + "\n", d.todos
def main():
ap = argparse.ArgumentParser(description="Render an ML scoping document from a JSON spec.", epilog=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("spec", nargs="?", help="JSON spec, or - for stdin")
ap.add_argument("-o", "--output")
ap.add_argument("--template", action="store_true")
ap.add_argument("--strict", action="store_true", help="exit 1 if any TODO remains")
a = ap.parse_args()
if a.template:
print(json.dumps(TEMPLATE, indent=2))
return 0
if not a.spec:
print("error: give a spec file (see --template)", file=sys.stderr)
return 2
try:
spec = json.loads(sys.stdin.read() if a.spec == "-" else open(a.spec, encoding="utf-8").read())
if not isinstance(spec, dict):
raise ValueError("spec must be a JSON object")
except (OSError, ValueError) as e:
print(f"error: {e}", file=sys.stderr)
return 2
text, todos = render(spec)
if a.output:
with open(a.output, "w", encoding="utf-8") as f:
f.write(text)
else:
sys.stdout.write(text)
if todos:
print(f"{len(todos)} TODO(s): " + "; ".join(todos), file=sys.stderr)
return 1 if todos and a.strict else 0
if __name__ == "__main__":
sys.exit(main())
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.