AI-driven credit platform for a regulated financial intermediary
A two-stage credit scoring engine with structural explainability, fully on-premise AI infrastructure, multi-layer encryption, and an audit trail that meets central-bank supervisory requirements.
The regulatory perimeter imposes three non-negotiable requirements: explainability of every automated credit decision, protection of financial and identity data, and full traceability that can withstand supervisory inspection.
Technical traits
On-premise perimeter with EU data residency and multi-layer encryption. SPID L2 → TLS 1.3 → JWT RS256 gateway → Conversational agent (on-prem LLM, state machine) → Document pipeline (dual OCR) → Scoring engine (GBT + calibration + SHAP) → Append-only event log. PII encrypted with AES-256-GCM. On-premise Vault. mTLS between services. Four separate VLANs.
Two-stage scoring with calibration
A tree-model ensemble followed by isotonic calibration that aligns predicted probabilities with observed default frequencies.
Explicit confidence band
Computed over a bootstrap ensemble, it widens as data grows sparse. Uncertainty is made visible rather than hidden behind a point estimate.
Supervised conversational agent
An on-premise LLM embedded in a finite state machine: every transition follows a deterministic path. The model handles linguistic variability — it does not govern the flow.
Append-only event log
Every state transition is an immutable event. A correction is an event that supersedes the previous one without deleting it. The audit trail is presentable with no additional processing.
Regulatory compliance is structural: it derives from the system architecture, not from third-party contractual guarantees. The system passes supervisory inspections by presenting an immediately queryable audit trail, with no additional processing required.