Cash where it is needed, not everywhere
Idle cash in branches and ATMs is capital earning nothing, and stock-outs are a service failure. Forecasting withdrawals per location turns replenishment routes and holdings into a solvable trade-off.
Decisions that survive an audit.
Optimization and forecasting for institutions where every decision has to be explainable after the fact. A model that cannot show its reasoning is not usable here, whatever its accuracy.
Three problems worth modelling
Collections effort aimed by expected recovery
Treating every delinquent account the same spends the most expensive channel on the least recoverable balances. Ranking by expected value directs contact where it pays.
Models a regulator can follow
Accuracy alone does not clear model risk review. Optimization gives an auditable objective and explicit constraints. The reasoning is the artefact, not a post-hoc explanation of a black box.
What we build
- Cash demand forecasting and ATM/branch replenishment optimization
- Collections prioritisation and channel allocation by expected recovery
- Branch and service-network location and capacity modelling
- Portfolio and asset-liability optimization under regulatory constraints
- Operations workforce and capacity planning against arrival patterns
- Document intelligence for KYC, contracts, and reporting, with audit trails
Methods: MILP, conic optimization, survival models, time-series, RAG with citations
Outcomes
- Less idle capital: cash held where withdrawals actually occur
- Higher recovery per contact: effort ranked by expected value, not age
- Defensible models: an explicit objective and constraints, not a black box