A Tier-1 financial institution needed to modernize its collections platform with production-grade machine learning that could handle 3.5 million monthly accounts. Within two quarters, a custom ensemble model delivered 3.2× revenue lift and an AI capability the bank can replicate across other functions.
North·Bank's collections operations processed 3.5 million accounts every month across multiple lending products. The team operated on a 20-year-old rules-based engine that scored accounts using static heuristics — last-payment date, balance band, days past due, and a handful of behavioral flags maintained by hand.
The model worked, but it was leaving real money on the table. Accounts that would self-cure were getting expensive outreach. Accounts likely to charge off weren't being prioritized fast enough. And the operations team couldn't experiment — every rule change required six weeks of risk and compliance review.
Leadership had an internal mandate to bring machine learning into collections, but two previous attempts had stalled. The team had built models that worked in notebooks but couldn't get them into production. Compliance review was a maze. The data infrastructure wasn't ready. Nobody had successfully bridged the gap between a data science prototype and a production system that risk, compliance, and operations all trusted.
We started with the operating reality, not the model. Before any data science work, we spent two weeks with the collections operations team and the risk function to understand what a deployable model actually needed to look like:
We built the data pipeline first. North·Bank had eight years of historical collections data spread across mainframe extracts, an aging Oracle warehouse, and several operational systems. We engineered a feature store on the bank's existing Databricks footprint, with 247 candidate features covering payment behavior, balance velocity, account tenure, and product mix.
We trained an ensemble of three models — a gradient-boosted classifier for charge-off risk, a regression model for expected recoverable balance, and a survival model for time-to-cure. The combined output produced an action recommendation per account that prioritized expected value, not just probability.
The bigger investment in this phase was the evaluation framework. We worked with the model risk management team to build an offline backtesting engine that could compare any model variant against the existing rules-based system across a representative population of historical accounts.
Production deployment used a champion-challenger pattern. The new model ran in shadow mode for four weeks on 100% of accounts, producing recommendations that were logged but not acted on. We compared shadow recommendations against actual collections outcomes and tuned thresholds before going live.
Once live, the model ran in active mode for 5% of accounts initially, with the population expanding weekly as the operations team gained confidence. By week 22, the model was driving prioritization for 100% of in-scope accounts, with humans always able to override individual recommendations.
"What made this different from our previous AI efforts wasn't the model — it was the way Alpine Quantum brought risk and compliance into the design process from day one. The model that shipped was the same one we showed risk in week four. That's how it cleared review in two months instead of two quarters."
Beyond the headline numbers, the engagement left North·Bank with three things that compound over time:
An AI capability they can extend. The feature store, evaluation framework, and deployment infrastructure are reusable across other risk and decisioning use cases. The team is already piloting a similar approach for credit line management.
A working relationship between data science and risk. The model risk management team now reviews model changes in days rather than weeks. The collaboration patterns built during this project are being applied to every new ML initiative.
Internal engineers who can operate it. Two of North·Bank's engineers worked alongside Alpine Quantum throughout the build. They now own day-to-day operations, model retraining, and the roadmap for what comes next.
Tell us what you're trying to solve. A senior AI engineer will reply within one business day.