Evidence, not adjectives
Eight cases on where money gets stuck, five of them with a model scoring live on the page. Every figure is either hers as reported, computed from public data, or labelled as a reconstruction.
The invoice that comes back
One invoice in six on a real wholesaler's ledger is later credited back. Every cancellation tied to the invoice it reverses (93% of the value), a model that predicts which invoices will come back at the moment they are raised (AUC 0.74 vs 0.68 for history alone), and a check-before-dispatch list with a cut-off.
When will this invoice be paid?
A payment-date model on 48,839 real B2B invoices: days-after-due with an honest band, a flag for the ones that slip past a week (AUC 0.87), and a $323M open book rolled into a weekly cash forecast — built with a no-lookahead protocol and judged against the collector's own number.
Who pays late?
A repayment-propensity model on 30,000 real accounts — six months of payment history in, a ranked chase list with an economic cut-off out. AUC 0.78, calibrated, fair across groups, exported to plain JavaScript and scoring live on the page.
The discount nobody approved
A medical-equipment procurement platform quoting hospitals by hand, losing money before the invoice existed. A pocket-price waterfall, a 3.1× price band for the same product, a corridor per category — and quote-to-order conversion up 27%.
The letter that must never be wrong
Inside a UK energy supplier's arrears path: the logic that decides which of hundreds of thousands of customers in debt gets which letter — and which must never get one. Four scripts, 8.5M-row tables, a tie-out that refuses to publish, zero escalations.
Six tiles, no answer
India's national delayed-payments dashboard tracks ₹55,244 crore owed to small businesses — as six coloured tiles of counts. Rebuilt on its own eleven reports: ₹4,769 crore has waited over a year, the headline froze last October, and the tables don't agree with the front page.
Where the money got stuck
A B2B startup delivering on time and getting paid late. The leak was the ledger, not the customers — on-time collections from 62% to 90%, discrepancies down 30%.
Fifty accounts, one ladder
A logistics firm where one in four clients paid on time. A six-rung reminder ladder, a tracker and two kinds of flexibility — payment adherence to 95%, receivables down 20%.
Why passengers walk away
A dissatisfaction model on 129,880 real journeys — AUC 0.99, calibrated, audited for a survey artefact — with a lever board: +1 on wifi removes ten walk-aways in a hundred; forty minutes less delay removes under one. Scores live on the page.
One DREQ, rebuilt on Databricks
A production client data-request workflow carried from Hadoop/Ambari into Databricks — SQL, PySpark, Delta tables, catalog discipline.
Experiments that moved sales
A/B tests and statistical analysis for campaigns and product features; Power BI dashboards leadership actually opened.
Hours returned to humans
Nine analyst-hours of weekly export-clean-paste automated down to forty minutes — with schema guards that fail loudly. The Seal in its purest form, and the daily job the collections work in case 01 ran on.