Architecting the AI Bridge: Integrating LLMs and DMN With Drools, Sp... Tim Wuthenow & Alex Porcelli

## Overview Tim Wuthenow and Alex Porcelli present a hybrid AI architecture that combines Large Language Models (LLMs) with Symbolic AI to overcome the limitations of each approach. Using a healthcare scenario—a patient with stage 4 cancer on a remote island with no hospital beds available—they illustrate why LLMs alone are insufficient: these models provide statistically correct but situationally inappropriate answers because they lack deterministic, context-aware decision logic tailored to specific business requirements.

## Key Topics The speakers argue that while LLMs excel at processing unstructured data and complex inputs, they are inherently non-deterministic and trained on static data. Conversely, Symbolic AI—specifically through Decision Model and Notation (DMN) and the Drools rule engine—provides deterministic, auditable, and business-friendly decision-making. The solution is a hybrid bridge that leverages both technologies via the MCP (Model Context Protocol).

## Technical Details **DMN (Decision Model and Notation):** An open standard that is graphically visualizable, human-readable, and executable. Decisions are modeled using decision tables (similar to spreadsheets with conditions and outcomes). DMN is non-developer-friendly, allowing business analysts and domain experts to define and modify rules. It is self-documented and LLMs are already trained to explain DMN logic.

**Drools:** An open-source rule engine maintained by the Apache Foundation that processes DMN models. It has been used for approximately 20 years.

**MCP Integration:** A Spring Boot application serves as an MCP server that exposes a catalog of available DMN decisions, along with their input and output schemas. This enables LLMs to discover, query, and execute decisions dynamically through a bidirectional interaction layer.

**Implementation:** A straightforward Spring Boot Maven project with DMN files in the resources folder and Drools as a dependency. The system provides REST endpoints, Java APIs, and native APIs alongside the MCP interface. In the demo, they illustrated querying available policies, understanding input requirements (e.g., "Is patient a cancer patient?", "Hospital beds available?"), executing decisions, and dynamically adding new rules (e.g., stroke-patient routing) without recompiling application code.

## Takeaways - **Hybrid AI is Essential:** Combining statistical (LLM) and deterministic (Symbolic AI) models produces robust, auditable, and contextually appropriate decisions. - **Auditability and Compliance:** DMN's deterministic nature and traceability are critical for regulated industries (healthcare, finance, insurance) where decisions must be replayed and explained to regulators. - **Instant Adaptation:** Changes to DMN rules are immediately available to LLMs through the MCP protocol, enabling real-time policy updates (e.g., Friday evening pricing changes). - **Business Ownership of Rules:** Domain experts can modify business logic without developer intervention, accelerating decision-making cycles. - **Open Standards Matter:** DMN's status as an open standard ensures vendor independence and LLM familiarity.