Financial institutions face a conflict between AML detection needs and data sovereignty. Cloud-based LLMs often struggle with the non-linear topologies of financial crime. This session presents a paradigm shift: an "AML Detective" agent for native graph reasoning on the edge, built with a 100% open-source stack.

We detail fine-tuning a 30B Mixture-of-Experts (MoE) model on the IBM AML Synthetic dataset. We show how Polars manages high-throughput transactions and NetworkX enables deterministic structural analysis. We explain our pivot from Reinforcement Learning (GRPO) to stable, process-driven Direct Preference Optimization (DPO) using Unsloth for memory-efficient training on local hardware. Using MLflow for trajectory tracking, our hybrid-judge system increased detection success by 10% while reducing investigation steps in half and ensuring sensitive data remains local. Attendees will learn to orchestrate these tools to build privacy-compliant, metacognitive agents bridging graph theory and generative AI.