Collections Optimizer

Daily collections plan within fair-collection rules. From today's overdue accounts (account_id, balance, dpd, promises, contact counts, consents, stop flags) and optional payment history (paid, contacted, channel, recovered), contact log, monthly snapshots, agents and test results it returns: propensity to pay (L2 logistic, weight-of-evidence driver report, held-out AUC); self-cure chance and contact uplift (two models, contacted vs not, held-out Qini; rule of thumb at cold start); a call queue ranked by expected extra recovery per call within agent capacity, with the best permitted channel, call window learned from reach rates, tone tier, why, and rule checks + id on every recommendation; self-cure, deferred (earliest permitted date) and do-not-contact lists; balance-first / oldest-first comparison; Markov roll rates, forecasts and eventual write-off chance; Kaplan-Meier time to cure; Hungarian agent allocation by language, skill and success; champion-challenger split and z-test. Hard limits: India 08:00-19:00, US 08:00-21:00 and Reg F 7-in-7 and 7-day wait, cease/dispute/bankruptcy/deceased/legal/hardship never queued; policy can only tighten. Decision cards first. Up to 5,000 accounts; larger: collections_optimizer_get_engine.

For books over 5,000 accounts or 20,000 rows of history, contact log or snapshots, or any such file already in your code sandbox. Returns a complete numpy-only Python script with the same engine as collections_optimizer; it reads accounts, history, contact log, snapshots, agents and test results, prints the decision cards, the queue and the do-not-contact list, and writes everything to collections_plan.json. No debtor data reaches this server on this path.