In the Netherlands, independent mortgage advisors (hypotheekadviseurs) face a fragmented landscape when placing complex or non-standard client files. While standard cases easily fit the major banks, edge cases—such as ZZP/entrepreneurs with short trading history, foreign income, non-standard employment contracts, ground leases (erfpacht), or specific energy-saving budgets—require advisors to manually scour through dozens of lengthy, frequently updated acceptance manuals (acceptatiegidsen, productvoorwaarden, and NHG criteria) from providers like Rabobank, ING, ABN AMRO, Florius, Munt Hypotheken, and Obvion.

In this challenge, students will design and build an AI-driven assistant that ingests publicly available acceptance manuals, policy PDFs, and regulatory frameworks (e.g., NHG / Nibud / TRHK norms) from Dutch mortgage lenders. The core technical challenges involve:

• Complex Document Parsing: Accurately extracting tables, rule hierarchies, exception clauses, and definitions from diverse PDF layouts.

• Domain-Specific Retrieval (RAG): Implementing hybrid search, vector embeddings, reranking, or agentic retrieval architectures capable of distinguishing subtle policy nuances across multiple lenders.

• Accuracy & Hallucination Mitigation: Ensuring the system strictly cites the exact page and section of the relevant policy guide, recognizing when a rule is missing or requires manual underwriter review rather than inventing answers.

1. Working Proof-of-Concept / Prototype: A functional RAG pipeline (e.g., chat interface, API, or MCP connector) that indexes public documentation from at least 3–5 Dutch mortgage providers plus NHG guidelines.

2. Standardized Evaluation Benchmark: A curated gold-standard test set of realistic broker queries (e.g., maximum age thresholds, probation periods, overseas income treatment, developer guarantees) paired with verified ground-truth answers.

3. Comparative Analysis & Research Report: A systematic comparison of different LLM architectures, chunking/retrieval strategies (e.g., hybrid vs. dense vector search, agentic search vs. standard RAG), evaluating response accuracy, citation fidelity, and latency.