SMB Growth Intelligence — live pipeline run

This page is generated, not written. The workflow runs make all on a clean runner, then renders whatever reports/*.json the pipeline produced. If the pipeline breaks or a quality assertion fails, that shows up here. Last run 2026-07-24 23:07 UTC · commit 445ddf3 · workflow run

Pipeline run

Ten SQL models across three layers — staging, intermediate, marts — executed in dependency order. Bars show relative execution time; the figure is in the next column.

Model	Layer	Rows	Seconds	Status
stg_invoice_lines	staging	392,690	0.5414s	✓ success
stg_invoice_lines_rejected	staging	134,737	0.1708s	✓ success
stg_products	staging	4,070	0.0212s	✓ success
stg_returns	staging	9,288	0.0194s	✓ success
int_orders	intermediate	18,532	0.0417s	✓ success
mart_cohort_retention	marts	91	0.0127s	✓ success
mart_country_revenue	marts	37	0.0084s	✓ success
mart_customer_features	marts	4,338	0.0292s	✓ success
mart_product_performance	marts	3,665	0.0557s	✓ success
mart_rfm_segments	marts	4,338	0.0090s	✓ success

541,909 raw line items in, 392,690 loaded, 14,482 duplicates removed and 134,737 quarantined (mostly guest checkouts with no customer id) — the difference is accounted for, not dropped.

Point-in-time feature store

8 features computed as of each cutoff date, so a training row only ever sees data that existed at that moment. The label is repeat_purchase_next_90d. ✓ leakage check passed

As-of cutoff	Rows	Positives	Base rate
2011-03-31	2,134	1,157	54.22%
2011-06-30	2,960	1,466	49.53%
2011-08-31	3,317	1,864	56.20%

Repeat-purchase model

Predicts whether a customer orders again within 90 days. Trained on cutoffs 2011-03-31, 2011-06-30 and validated on 2011-08-31 — 5,094 training rows, 3,317 validation rows at a 56.20% base rate. The baseline is the obvious heuristic: rank customers by how recently they last ordered.

Metric	Model	Baseline	Δ
ROC-AUC	0.7305	0.6838	+0.0467
PR-AUC	0.7949	0.7040	+0.0909
Top-decile lift	1.683×	1.360×	+0.323

Read this honestly: the model beats a recency-only baseline on every metric — ROC-AUC, PR-AUC and top-decile lift — on real, unseen customers. The top predictor is prior order frequency, which is exactly what you'd expect to drive repeat purchase.

Feature weights

Signed model coefficients. Direction is shown by which side of the axis the bar falls as well as by colour, and every bar carries its value.

▲ increases propensity ▼ decreases propensity

Feature	Weight	Value
orders_last_180d		▲ +1.3942
units_last_90d		▲ +1.0106
avg_line_value_90d		▼ -0.9569
distinct_products_90d		▲ +0.2320
orders_last_90d		▲ +0.2188
revenue_last_90d		▼ -0.1242
days_since_last_order		▼ -0.1010
revenue_last_180d		▲ +0.0354

Data-quality assertions

11 YAML-declared assertions run against the warehouse after every pipeline execution, written to an audit table and scored. Critical failures fail the build.

Check	Type	Severity	Observed	Threshold	Status
sales_volume	row_count_min	critical	3.927e+05	1e+05	✓ pass
sales_freshness	freshness	critical	1	2	✓ pass
customer_id_not_null	not_null	critical	0	0	✓ pass
unit_price_positive	min_value	high	0.001	0	✓ pass
quantity_positive	min_value	high	1	1	✓ pass
orders_reconcile_lines	reconcile_count	high	1.853e+04	1.853e+04	✓ pass
customer_features_unique	unique	high	0	0	✓ pass
monetary_non_negative	min_value	high	3.75	0	✓ pass
rfm_segments_not_orphaned	relationship	medium	0	0	✓ pass
propensity_scores_not_orphaned	relationship	medium	0	0	✓ pass
unattributable_rows_within_slo	quarantine_rate	medium	0.2486	0.3	✓ pass

How it is put together

01

Ingest Real UCI Online Retail invoices; guest checkouts and cancellations quarantined at the boundary.

→

02

Model Layered SQL: staging → intermediate → marts, run in dependency order.

→

03

Features Config-driven point-in-time store, backfilled per cutoff.

→

04

Train Walk-forward fit, scored back into the warehouse as a mart.

→

05

Assert Quality checks to an audit table; critical failures break the build.

It runs locally on DuckDB with no cloud account and no cost, and the same modelling logic ports to BigQuery and Dataform — see /bigquery.

make setup make all pytest -q

392,690 sale lines to scored customers in 34.49s.

Pipeline success

Records quarantined