The Problem

Insurance loan applications involve dense regulatory documents — FSCA requirements, NCA compliance, product-specific policy terms — that must be verified for every application. Manual compliance checking is slow, expensive, and inconsistently applied. Errors create regulatory liability.

INSURELOANSA automates this compliance layer entirely.

What It Does

The system ingests an insurance loan application and runs it through a RAG pipeline that has ingested the full regulatory corpus:

FSCA regulatory frameworks
National Credit Act requirements
Product-specific policy schedules
Internal underwriting guidelines

For each application, it produces:

Compliance Assessment — pass/fail against every applicable regulatory requirement, with the specific clause cited for each finding
Risk Flags — items requiring human review, ranked by severity
Structured Report — formatted output ready for the underwriter's workflow

Architecture

Application Input (API or form)
        ↓
FastAPI Processing Service
        ↓
Document Parser (extract key fields)
        ↓
Compliance RAG Pipeline
    - Query ChromaDB (regulatory corpus)
    - Retrieve relevant requirements
    - Claude generates structured assessment
        ↓
Risk Scoring Engine
        ↓
Report Generation (PDF + JSON)
        ↓
Underwriter Dashboard

Key Design Decision: Structured Output

Every compliance finding must be traceable to a specific regulation. I implemented a structured output schema using Anthropic's tool use API — Claude is forced to populate a typed schema for every finding, including the source document, section reference, and verbatim requirement text. No unstructured opinions, no hallucinated citations.

Performance

Average processing time per application: under 90 seconds
Regulatory corpus: 847 documents, ~2.3M tokens embedded
False positive rate on compliance flags: less than 8% (validated against manual review)

What I Learned

Insurance compliance requires extreme precision on citations. The first version used standard RAG with summarisation — regulators couldn't trace the AI findings back to source documents. Switching to verbatim extraction with forced structured output resolved the trust issue entirely.