Inspiration

very year, Malaysian banks receive over RM47.3 Billion in property finance applications. Yet, processing a single mortgage application takes a staggering 15 to 22 days. Borrowers (especially in the B40 and M40 segments), suffer under long waiting times and subjective human biases, while bank compliance desks remain blind to relational "fraud-for-profit" networks

Legacy systems evaluate each loan file in strict tabular isolation, making them completely blind to coordinated fraud rings where corrupt appraisers, employers, and brokers share contact details and falsify records across multiple different applications

MLTF (Mortgage Loan for The Future) is an automated framework that integrates a Large Language Model (LLM)-based multi-agent system with a dynamic graph database. It is designed to slash operational processing latency from weeks down to hours. By replacing legacy OCR with specialized AI agents, the system extracts context-aware document data and maps it into a relational network to uncover hidden, multi-party fraud patterns while routing borderline applications to human underwriters via a Human-in-the-Loop (HITL) dashboard.

How we built it

Architecture: We implemented a Supervisor-Worker Agentic AI Pattern natively in Java 21 and Spring Boot 3.x, orchestrated via the Google Agent Development Kit (ADK).
Document Extraction: We utilized Gemini 2.0 Flash to process raw PDFs through multimodal reasoning, bypassing traditional OCR limitations.
Fraud Detection: A specialized Fraud Anomaly Worker leverages Gemini 1.5 Pro to execute multi-hop network traversals in Google Cloud Spanner Graph using ISO Graph Query Language (GQL).
Compliance: An asynchronous Audit Worker uses Gemini 2.0 Flash Lite to generate human-readable, PDPA-compliant audit logs stored as immutable JSON.
Deployment: The client interface was built using Angular (v17+) and Angular Material, deployed as serverless microservices on Google Cloud Run.

Challenges we ran into

During the deployment, we facing issue about hallucination, when the model start imagination a new data that is not exist in the dataset. To prevent that we utilizing gatekeeper that guardrail the model reasoning to prevent them from hallucinations.

Accomplishments that we're proud of

Successfully engineering a system capable of overcoming OCR context blindness and detecting coordinated, multi-party fraud rings.
Leveraging Spanner Graph's Index-Free Adjacency to navigate direct memory pointers, allowing complex relational queries to execute in milliseconds.
Creating a powerful "force multiplier" for back-office banking personnel that simultaneously removes subjective human bias from underwriting.

What we learned

We learned by combining 2 advanced technology: Agentic AI and Graph Database, we can create state of the art solution that novels and robust to tackle the real world problem on Mortgage Loan Origination System, and help the Nation to be better place to the world.

What's next for MLTF: Mortgage Loan for The Future

Scaling the prototype to capture the entire complexity of core-banking systems and data interconnections in major international financial institutions.
Expanding the multi-agent workflow to include other loan processing activities, such as loan product management, collateral valuation, and legal processes