A three-layer architecture designed for regulated credit environments
I structured the platform as three deliberately separated components: a signal
ingestion and feature engineering layer, a model inference engine, and a
lender-controlled policy layer. Each component has a defined boundary and a specific
responsibility. The ingestion layer normalises and validates incoming financial
signals before any of them reach the model. The inference engine processes clean
features and produces a recommendation alongside a confidence score and a ranked
signal explanation. The policy layer applies the lender's configured rules
before any output is surfaced. This separation is not an engineering preference — it
is a compliance requirement. Australian responsible lending obligations require that
a human institution sets the decision rules. The platform architecture reflects that
requirement structurally, not just in documentation. When a lender's compliance
team asks how a recommendation was produced, the answer is traceable through three
discrete, logged components — each of which can be audited independently.
The signal ingestion layer connects to the lender's existing data
environment without requiring new data collection infrastructure. Repayment
history, utilisation patterns, income indicators, and account activity are read
on a rolling basis, normalised against defined quality thresholds, and passed to
the feature engineering process only when they meet completeness standards.
The inference engine applies a calibrated model to the engineered features and
produces three outputs simultaneously: a recommended credit limit adjustment, a
confidence indicator, and a ranked explanation of the signals that most
influenced the recommendation. These three outputs are generated together — the
explanation is not reconstructed after the fact.
The policy layer sits between the model output and the recommendation delivered
to the lender. It applies adjustment bands, borrower segment filters, minimum
confidence thresholds, and any institution-specific exclusion criteria
configured by the lender. Recommendations that do not pass the policy check are
flagged for manual review with a structured explanation — they are never
silently suppressed.
Every component logs its outputs continuously. The ingestion layer records the
signals received and their quality check results. The inference engine logs the
model version, the features used, and the outputs produced. The policy layer
logs which rules were applied and what the outcome was. This creates a complete,
retrievable record for every recommendation the platform generates.