Apex Lending Intelligence Suite
Sub-100 ms credit decisions with an auditable ensemble
A non-bank lender needed a loan-origination platform that could decide at peak application volume without sacrificing model quality or the audit trail regulators expect.
The challenge
- Decisions had to return in under 100 milliseconds at peak.
- Features came from bureau data, bank statements and alternative sources with different freshness.
- Every repayment event had to be stored durably at high write rates.
What we built
Ensemble over gradient boosting
XGBoost and LightGBM models combined, with monotonic constraints and reason codes so each decision is explainable.
A decisioning service, not a notebook
A Python/FastAPI service serves the ensemble with versioned models and feature validation on every request.
Storage matched to workload
Cassandra stores the high-write transaction and repayment-event ledger; Redis fronts the scoring API for hot features.
Automation around the decision
n8n orchestrates document collection, KYC verification calls and underwriter notifications across the CRM and e-signature tools.
What we learned
- Feature freshness mattered more than model complexity for accuracy at decision time.
- Putting the ledger on Cassandra removed the write bottleneck that a relational store had hit.
- Workflow automation around the model delivered as much operational value as the model itself.
Want something like this?
Most of these projects started as a two-page brief. Send us yours.